Home >
Documents >
Deconvolution and regularization with Toeplitz matrices · 2014-11-17 · Deconvolution and...

Share this document with a friend

Embed Size (px)

of 56
/56

Transcript

Numerical Algorithms 29: 323–378, 2002. 2002 Kluwer Academic Publishers. Printed in the Netherlands.

Deconvolution and regularization with Toeplitz matrices

Per Christian Hansen

Department of Mathematical Modelling, Technical University of Denmark, Building 321,DK-2800 Lyngby, DenmarkE-mail: [email protected]

Received 2 August 2000; revised 5 November 2001

By deconvolution we mean the solution of a linear first-kind integral equation with aconvolution-type kernel, i.e., a kernel that depends only on the difference between the two in-dependent variables. Deconvolution problems are special cases of linear first-kind Fredholmintegral equations, whose treatment requires the use of regularization methods. The corre-sponding computational problem takes the form of structured matrix problem with a Toeplitzor block Toeplitz coefficient matrix. The aim of this paper is to present a tutorial survey ofnumerical algorithms for the practical treatment of these discretized deconvolution problems,with emphasis on methods that take the special structure of the matrix into account. Whereverpossible, analogies to classical DFT-based deconvolution problems are drawn. Among otherthings, we present direct methods for regularization with Toeplitz matrices, and we show howToeplitz matrix–vector products are computed by means of FFT, being useful in iterative meth-ods. We also introduce the Kronecker product and show how it is used in the discretizationand solution of 2-D deconvolution problems whose variables separate.

Keywords: deconvolution, regularization, Toeplitz matrix, Kronecker product, SVD analysis,image deblurring

AMS subject classification: 65F30

1. Introduction to deconvolution problems

The main purpose of this survey paper is to present modern computational methodsfor treating linear deconvolution problems, along with some of the underlying theory.Discretization of these problems leads to structured matrix problems with a Toeplitzor block Toeplitz coefficient matrix. The main focus is on numerical aspects, and weillustrate how mathematics helps to derive efficient numerical deconvolution algorithmsthat exploit the Toeplitz structure of the matrix problems. Our goal is to demonstrate thatfast deconvolution can be performed without the artificial periodization of the data thatis common in many popular FFT-based algorithms. The present paper can be viewed asan addendum to the monograph [20] where Toeplitz regularization algorithms are onlymentioned briefly.

324 P.C. Hansen / Deconvolution and regularization

The term “deconvolution” seems to have slightly different meanings in differentcommunities – to some, it is strongly connected to the operation of convolution in dig-ital signal processing, while to others it denotes a broader class of problems. In theremainder of this introduction, we shall briefly describe both meaning of the term.

Before we present the algorithms, we make a detour in section 2 into the world ofinverse problems and, in particular, Fredholm integral equations of the first kind. Thereason for this is that deconvolution problems are special cases of these integral equa-tions. Extensive amounts of both theory and algorithms have been developed for thesegeneral problems, and our detour will provide us with important insight and therefore astronger background for solving deconvolution problems numerically.

Having thus set the stage, we turn to the numerical methods in sections 3–6. Westart in section 3 with a general discussion of the discretization and numerical treatmentof first-kind integral equations, and we introduce the singular value decomposition whichis perhaps the most powerful “tool” for understanding the difficulties inherent in theseintegral equations. Next, in section 4 we discuss general regularization algorithms forcomputing stabilized solutions to the discretized linear systems, with no emphasis onmatrix structure.

In the remaining sections we turn to numerical methods that exploit matrix struc-ture. In section 5 we discuss various direct and iterative methods for one-dimensionalproblems, and show how we can compute the numerical solutions accurately and ef-ficiently. The emphasis here is on numerically stable methods based on orthogonaltransformations and FFT, and we also consider the use of stabilized hyperbolic rota-tions. Finally, in section 6 we turn to 2-dimensional problems, and here we focus onproblems in which the variables separate, allowing us to also develop very efficient nu-merical schemes for these problems. The emphasis here is on Kronecker products, theirproperties, and their use in direct and iterative methods. In addition, we briefly discussthe use of the 2-D FFT algorithm.

Throughout the paper, the theory and the algorithms are illustrated with represen-tative problems from gravity surveying and image deblurring.

We do not attempt to survey the vast area of deconvolution and its applications inscience and engineering. The literature on this subject is rich, and presentations at alllevels of sophistication can be found. The updated collection of papers in [23] providesboth background material and surveys of applications, and is a good starting point.

Some acquaintance with numerical linear algebra is necessary; the most completereference is the book [15] that discusses all state-of-the-art techniques, classical as wellas modern. The monograph [5] contains a wealth of material on numerical least squaresmethods. We also wish to mention the recent monograph [24] devoted to theory andalgorithms for structured matrices, including a wealth of material on Toeplitz matrices.

In our presentation we shall frequently use the elementwise product (Hadamardproduct), denoted by �, of two conforming matrices or vectors. Similarly we use thesymbol � to denote the elementwise division of two conforming matrices or vectors.These operations are identical to the Matlab operations “.*” and “./”, respectively.

P.C. Hansen / Deconvolution and regularization 325

We shall also make reference to the Matlab package REGULARIZATION TOOLS

[19,21] from time to time. More details about the underlying theory that is briefly pre-sented in this paper can be found in the “deconvolution primer” [35], while a recentsurvey of numerical methods for general linear inverse problems is given in the mono-graph [20].

1.1. Deconvolution in digital signal processing

Given two discrete (digital) real signals f and h, both of infinite length, the convo-lution of f and h is defined as a new infinite-length signal g as follows

gi =∞∑

j=−∞fjhi−j , i ∈ Z.

Note that one (or both) signals may have finite length, in which case the above relationstill holds when the remaining elements are interpreted as zeros. The convolution is usedto represent many useful operations in digital signal processing. For example, if h is theinput signal then the output of a FIR filter with n nonzero filter coefficients f0, . . . , fn−1

is given by

gi =n−1∑j=0

fjhi−j , i ∈ Z.

Throughout, let ı denote the imaginary unit satisfying ı2 = −1. Then g(ω) =∑∞j=−∞ gje

−2πıjω is the Fourier transform of g, and the following important relationholds for the Fourier transforms of f , g, and h:

g(ω) = f (ω)h(ω), ω ∈ R.

If the two discrete signals f and h are both periodic with periodN , and representedby the sequences f0, f1, . . . , fN−1 and h0, h1, . . . , hN−1, then the convolution of f andh is another periodic signal g with period N , and the elements of g are defined as

gi =N−1∑j=0

fjh(i−j)mod N, i = 0, 1, . . . , N − 1.

Note that the subscript of h is taken modulo N . The discrete Fourier transform (DFT)of g is defined as the sequence

Gk ≡ 1

N

N−1∑j=0

gje−ı 2πjk/N, k = 0, 1, . . . , N − 1,

and it is convenient to denote the N-vector consisting of this sequence by DFT(g). Thenthe following relation holds for the DFT of the convolution of f and h

DFT(g) = DFT(f )� DFT(h), (1)

326 P.C. Hansen / Deconvolution and regularization

i.e., the vector DFT(g) equals the element-wise product of the two vectors DFT(f ) andDFT(h). Consequently, g = IDFT(DFT(f )�DFT(h)), where IDFT denotes the inverseDFT. Note that f and h can be interchanged without changing g. We remind that theDFT and the IDFT of a signal can always be computed efficiently by means of the fastFourier transform (FFT) in O(N log2N) operations. A very thorough discussion of theDFT (and the FFT) is presented in [7].

Deconvolution is then defined as the process of computing the signal f given theother two signals g and h. For periodic discrete signals, equation (1) leads to a simpleexpression for computing f , considered as an N-vector:

f = IDFT(DFT(g)� DFT(h)

). (2)

Again, the FFT can be used to perform these computations efficiently. For general dis-crete signals, the Fourier transform of the deconvolved signal f is formally given byf (ω) = g(ω)/h(ω), but there is no similar simple computational scheme for computingf – although the FFT and equation (2) are often used (and occasionally misused) forcomputing approximations to f efficiently; cf. section 5.5.

1.2. General deconvolution problems

Outside the field of signal processing, there are many computational problems thatresemble the classical deconvolution problem in signal processing, and these problemsare also often called deconvolution problems. Given two real functions f and h, thegeneral convolution operation takes the generic form

g(s) =∫ 1

0h(s − t)f (t) dt, 0 � s � 1, (3)

where we assume that both integration intervals have been transformed into the interval[0, 1]. Certain convolution problems involve the interval [0,∞), but in order to focusour presentation on the principles of deconvolution we omit these problems here.

The general problem of deconvolution is now to determine either f or h, given theother two quantities, and we shall always assume that h is the known function while wewant to compute f . Usually, the function h comes from a mathematical model of theunderlying problem, while g is available as measured data for a discrete set of s-values,i.e., we are given “noisy samples” of g at certain discrete abscissas s1, s2, . . . , sm:

gi = g(si)+ ei, i = 1, . . . , m.

Here, ei denotes the measurement noise associated with the ith data point. The noiseusually comes from some statistical distribution, which may be known or unknown.

One example of deconvolution is the following model problem in gravity surveying(presented as a geomagnetic prospecting problem in [35, p. 17]). Assume that a one-dimensional horizontal mass distribution lies at depth d below the surface – see figure 1for the geometry and the location of the s and t axes. From measurements of the verticalcomponent of the gravity field, denoted g(s), at the surface, we want to compute the mass

P.C. Hansen / Deconvolution and regularization 327

Figure 1. The geometry of the gravity surveying model problem. The measured signal g(s) is the verticalcomponent of the gravity field due to a 1-D mass distribution f (t) at depth d .

distribution, denoted f (t), along the t axis. The contribution to g from an infinitesimalpart dt of the mass distribution at t is given by

dg = sin θ

r2f (t) dt,

where the angle θ is shown in figure 1, and the distance between the two points on the sand t axes is given by r = √

d2 + (s − t)2. Using that sin θ = d/r, we get

sin θ

r2f (t) dt = d

(d2 + (s − t)2)3/2f (t) dt.

The total value of g(s) for any s is therefore

g(s) =∫ 1

0

d

(d2 + (s − t)2)3/2f (t) dt.

Thus, we arrive at a deconvolution problem for computing the desired quantity f withkernel given by h(s − t) = d(d2 + (s − t)2)−3/2.

Deconvolution problems also arise, e.g., in statistics, and we shall briefly describean example from astrometry. Assume that we are given a measured statistical distribu-tion g of stellar parallaxes (the parallax is a small angle from which the distance to thestar can be computed). This distribution is perturbed due to various measurement errors,and it can be shown that if f denotes the true distribution of parallaxes, then f and g arerelated by the equation∫ 1

0

1

σ√

2πexp

(−1

2

(s − tσ

)2)f (t) dt = g(s),

where σ is a parameter that characterizes the accuracy of the particular set of measure-ments. The above equation is obviously a deconvolution problem with kernel

h(s − t) = (σ√2π)−1

exp

(−1

2

(s − tσ

)2)

328 P.C. Hansen / Deconvolution and regularization

for computing the true distribution f of parallaxes. This problem is implemented as thefunction parallax in REGULARIZATION TOOLS.

The inverse heat equation [12] is yet another example of a deconvolution problem.Assume that we wish to determine the temperature f , as a function of time t , on oneside of a wall (which is inaccessible) given measurements g of the temperature on theother (accessible) side of the wall. This leads to a deconvolution problem of the genericform ∫ s

0h(s − t)f (t) dt = g(s) (4)

with kernel h given by

h(s − t) = (s − t)−3/2

2κ√π

exp

(− 1

4κ2(s − t)),

where the parameter κ describes the heat conduction properties of the wall. This modelproblem is implemented in REGULARIZATION TOOLS as the function heat.

2. First-kind Fredholm integral equations

At this stage, we begin our detour into the world of inverse problems and Fredholmintegral equations of the first kind. Inverse problems can often be characterized as prob-lems in which we wish to compute certain properties of the interior of a domain, givenmeasurements made from the outside along with a mathematical model of the relationbetween the interior and the measured data. Seismology is a classical example of aninverse problem, where the goal is to map the various layers of the earth, given measure-ments at the earth’s surface of seismic waves penetrating the layers and reflected by thelayers. Another classical example is computerized X-ray tomography, where we wishto compute images of, say, the human brain, given measurements of the damping of theX-rays through the brain.

It should be emphasized here that inverse problems do not always involve a strictlyinterior region. It is perhaps more precise to say that from measured data one wants toinfer about certain hidden data, unaccessible to be measured directly. The two examplesmentioned above are clearly in this class of problems. Another example is image deblur-ring: here, given a recorded (e.g., digitized) blurred image and a mathematical model forthe blurring process, the goal is to reconstruct the original sharp image as accurately aspossible. The Hubble Space Telescope is perhaps the most famous example of the needfor image deblurring [1].

All the above-mentioned inverse problems, as well as the deconvolution problemsfrom the previous section, can be formulated as Fredholm integral equations of the firstkind. The generic form of a first-kind Fredholm integral equation looks as follows∫ 1

0K(s, t)f (t) dt = g(s), 0 � s � 1, (5)

P.C. Hansen / Deconvolution and regularization 329

where the function K, called the kernel, is a known function of the two variables s and t ,and the right-hand side g is also known – or at least measured for discrete values of s –while f is the unknown function that we wish to compute. Note that the deconvolutionproblems from the previous section are merely special cases of (5) in which the kernelK(s, t) = h(s − t) depends solely on the difference s − t between the two variables.

We point out that the integration interval in equation (4) is from 0 to s; such anintegral equation is called a Volterra integral equation of the first kind. It can be consid-ered as a special Fredholm integral equation whose kernel is zero for t > s and it sharesthe same difficulties as the Fredholm equation.

2.1. Smoothing and inversion

As already mentioned, the advantage of working with the general formulation (5) isthat the underlying theory of first-kind Fredholm integral equations is so well developed.For example, it is well understood that the integration of f withK is a smoothing opera-tion that tends to dampen high-frequency components in f , such that the function g is asmoother function than f . Consequently, the inverse process, namely, that of computingf from g, can be expected to amplify the high-frequency components and it is thus a“desmoothing process,” so to speak.

The above statement is quantified by the following example. If we choose thefunction f to be given by

f (t) = sin(2πpt), p = 1, 2, . . . ,

such that the corresponding right-hand side g is given by

g(s) =∫ 1

0K(s, t) sin(2πpt) dt, p = 1, 2, . . . ,

then the Riemann–Lebesgue lemma1 states that

g→ 0 as p→∞.In other words, the higher the frequency of f , the more g is damped – and this is trueindependently of the kernel K; see figure 2 for a numerical example. Consequently, thereverse process, i.e., that of computing f from g, will amplify the high frequencies.

If all data were unperturbed, and all computations could be done in infinite pre-cision, then the inversion process in getting from g to f would be perfectly possible.Unfortunately, this situation never arises in practice: measured data are always conta-minated by errors, and numerical computations always involve small but non-negligiblerounding errors. Due to the above result, small perturbations of the high-frequency com-ponents in g are transformed into large perturbations of f , and the higher the frequencythe larger the perturbation of f . Hence, the unavoidable data and rounding errors make

1 The Riemann–Lebesgue lemma can be formulated as follows: if the function ψ has limited total fluctua-tion in the interval (0, 1), then, as λ→∞,

∫ 10 ψ(θ) sin(λθ) dθ is O(λ−1).

330 P.C. Hansen / Deconvolution and regularization

Figure 2. Illustration of the Riemann–Lebesgue lemma for the gravity surveying problem, with f (t) =sin(2πpt), g(s) = ∫ 1

0 h(s − t)f (t)dt , and p = 1, 2, 4 and 8.

the practical inversion process a very unstable process indeed, and in practice it is impos-sible to compute or estimate f by means of a direct inversion. Some kind of stabilizationtechnique is needed in order to recover a reasonably accurate approximation to the de-sired solution f .

2.2. An example from signal processing

We can illustrate the above-mentioned difficulties by means of the deconvolutionformula (2) for periodic digital signals. If h corresponds to a smoothing operation, thenthe spectral components of h, i.e., the components of the vector DFT(h), are such thatthe higher the frequency, the smaller the corresponding element. Hence, the same is truefor the elements of the DFT of the exact right-hand side. Assume now that we are giventhe vector g = g + e of measured values, consisting of the exact data g contaminatedby additive white noise e with elements from a normal distribution with zero mean andstandard deviation σnoise. Then the DFT of g is given by

DFT( g ) = DFT(g)+ w,where all the elements in the vector w = DFT(e) have the same probability. Thus, theexpression for the DFT of the computed solution f becomes

P.C. Hansen / Deconvolution and regularization 331

Figure 3. Power spectra of the various signals involved in the low-pass filtering example.

DFT(f)=DFT( g )� DFT(h) = DFT(g)� DFT(h)+ w � DFT(h)

=DFT(f )+ w � DFT(h),

showing that the high-frequency components of f are perturbed the most, due to thedivision of the elements in w by the small elements in DFT(h).

We illustrate the above with a numerical example involving a low-pass filter withfilter coefficients 0.5, 1, 1, 1 and 0.5, applied to a short sequence of 512 samples ofa speech signal. Figure 3 shows various power spectra of the signals involved in thisexample. The noise e is generated in Matlab as e = 0.1*randn(512,1), and the powerspectrum of e is flat, as we expect from white noise. Notice how the high-frequencycomponents of the deconvolved signal are perturbed wildly, especially around the zerosof the low-pass filter! Clearly, “naive” inversion by means of equation (2) is useless.

3. Numerical treatment

Before we turn our attention to practical schemes for stabilizing the inversionprocess, it is convenient to discuss the numerical treatment of integral equations, as

332 P.C. Hansen / Deconvolution and regularization

well as some important concepts from numerical linear algebra associated with ill-conditioned matrices, most notably the singular value decomposition.

3.1. Discretization by quadrature rules

There is a variety of schemes available for discretization of integral equations,i.e., for turning the integral equation into a system of linear equations that can be solvednumerically in order to provide an approximate discrete solution to the integral equation.The most complete survey is probably [2]. Here, for pedagogical reasons, we limit ordiscussion to quadrature methods based on well-known quadrature rules. Recall that aquadrature rule for computing an approximation to a definite integral takes the followinggeneral form ∫ 1

0ϕ(t) dt �

n∑j=1

wjϕ(tj ),

where t1, . . . , tn are the abscissas for the particular quadrature rule, and w1, . . . , wn arethe corresponding weights. For example, for the midpoint rule (which is identical to thetrapezoidal rule for periodic functions), we have

tj = j − 0.5

n, wj = 1

n, j = 1, . . . , n,

while for Simpson’s rule (where n must be odd) the abscissas are given by

tj = j − 1

n− 1, j = 1, . . . , n,

and the weights are

(w1, w2, w3, w4, w5, . . . , wn−1, wn) = (c, 4c, 2c, 4c, 2c, . . . , 4c, c)with c = 1/(3(n − 1)). This formulation also covers more advanced rules such as theNewton–Cotes rules. Using a quadrature rule, we can approximate the integral in ourFredholm integral equation as follows∫ 1

0K(s, t)f (t) dt �

n∑j=1

wjK(s, tj )f (tj ) = ψ(s),

and we emphasize that the right-hand side ψ(s) in the above expression is still a functionof the variable s. Notice that we have replaced f with f because we introduce approxi-mation errors in the above expression and thus cannot expect to compute f exactly.

In order to arrive at a system of linear equations, we can now use collocation, i.e.,we require that the function ψ defined above equals the right-hand side g at given pointss1, . . . , sm:

ψ(si) = g(si), i = 1, . . . , m.

P.C. Hansen / Deconvolution and regularization 333

Here, the numbers g(si) are usually the measured values of the function g. Note that mneed not necessarily be equal to n, but to keep our exposition simple we shall assume sothroughout the paper, i.e., m = n. Inserting the expression for ψ(s) into the collocationscheme, we arrive at the following equations

n∑j=1

wjK(si, tj )f (tj ) = g(si), i, j = 1, . . . , n.

When we rewrite these equation in matrix notation, we obtain the systemw1K(s1, t1) w2K(s1, t2) . . . wnK(s1, tn)

w1K(s2, t1) w2K(s2, t2) . . . wnK(s2, tn)...

......

w1K(sn, t1) w2K(sn, t2) . . . wnK(sn, tn)

f (t1)

f (t2)...

f (tn)

=g(s1)

g(s2)...

g(sn)

or simply Ax = b, where A is an n× n matrix. The elements of A, b and x are given by

aij =wjK(si, tj )bi = g(si)xj = f (tj )

i, j = 1, . . . , n.

To illustrate the above scheme, we use the midpoint rule to discretize the gravitysurveying problem from section 1, with quadrature and collocations points equidistantlydistributed in the interval [0, 1] as si = ti = (i − 0.5)/n, i = 1, . . . , n. Thus, the matrixelements are given by

aij = d/n

(d2 + (si − tj )2)3/2 =n2d

(n2d2 + (i − j)2)3/2 , i, j = 1, . . . , n.

As the solution in our model problem, we choose f (t) = sin(πt)+0.5 sin(2πt), and theelements of the exact solution x thus consists of the sampled values of f at the abscissastj = (j−0.5)/n for j = 1, . . . , n. Finally, the right-hand side b is computed as b = Ax.Due to our choice of quadrature and collocation points we obtain a symmetric matrix,and the depth is chosen to be d = 0.25.

At this stage, we emphasize that in practice the right-hand side is usually a per-turbed version of this b. That is, we solve the system Ax = b, where b = b+ e, and thevector e represents the perturbation of the exact data.

The elements of the computed vector xnaive = A−1b are, in principle, approxima-tions to the desired solution, i.e., we compute “samples” of the function f at theabscissas t1, . . . , tn given by

f (tj ) = (xnaive)j , j = 1, . . . , n.

From a naive point of view one would think that these quantities are supposed to approx-imate the solution f at the same abscissas. But as we have already emphasized severaltimes, we cannot expect the naively computed solution to a first-kind Fredholm integralequation to be stable with respect to perturbations of then right-hand side, and hence we

334 P.C. Hansen / Deconvolution and regularization

Figure 4. “Naive solution” to the gravity surveying model problem.

should not expect the computed values f (tj ) to be good approximations to the quantitiesf (tj ) if we solve the system Ax = b by standard tools in numerical analysis.

Figure 4, which shows the “naive” solution xnaive to the gravity model problem,illustrates this fact – the solution has nothing in common with the exact solution, thenorm is not even of the correct order of magnitude. We postpone the discussion of howto stabilize the computed solutions to the next section, and turn our attention to tools-of-the-trade in numerical linear algebra.

3.2. The singular value decomposition

One of the most versatile tools in linear algebra is probably the singular value de-composition (SVD) of a matrix [15], and we shall make heavy use of this decompositionthroughout the paper. The SVD is defined for any rectangular matrix, but we limit ourdiscussion to square n× n matrices, for which the SVD takes the form

A = U(V T =n∑i=1

uiσivTi .

The two matrices U and V consist of the left and right singular vectors,

U = (u1, . . . un), V = (v1, . . . , vn),

and both matrices are orthogonal, i.e., UTU = V TV = I . This implies that the leftand right singular vectors are orthonormal, uT

i uj = vTi vj = δij . The middle matrix

( = diag(σ1, . . . , σn) is a diagonal matrix whose diagonal elements σi are the singularvalues of A. They are nonnegative and ordered in non-increasing order, i.e.,

σ1 � σ2 � · · · � σn � 0.

P.C. Hansen / Deconvolution and regularization 335

It can be shown that the number of nonzero singular values is equal to the rank of A andthat the condition number of A is cond(A) = σ1/σn. Software for computing the SVDis available in all modern software packages.

If A is a symmetric matrix, A = AT, then the SVD of A is related to the eigenvaluedecomposition A = W/WT with W = (w1, . . . , wn) and / = diag(λ1, . . . , λn) asfollows

(ui, σi, vi) ={(wi, λi, wi) for λi � 0,(wi,−λi,−wi) for λi < 0,

(6)

and this relation can be used to simplify the computation of the SVD for symmet-ric matrices. In particular, if A is symmetric and positive definite then (ui, σi, vi) =(wi, λi, wi) for all i.

The singular values and vectors satisfy a number of important relations, and thefollowing is the most important:

Avi = σiui, ‖Avi‖2 = σi, i = 1, . . . , n. (7)

This relation leads to a simple expression for the solution in terms of the SVD. First notethat we can expand b and x in terms of the left and the right singular vectors ui and vi ,respectively, as

b =n∑i=1

(uTi b)ui, x =

n∑i=1

(vTi x)vi.

The latter relation, combined with (7), leads to the expression

Ax =n∑i=1

σi(vTi x)ui,

and equating the expressions for b and Ax, we arrive at the relations (uTi b) = σi(v

Ti x)

for i = 1, . . . , n. Hence, the “naive” solution to Ax = b can be written as

xnaive =n∑i=1

uTi b

σivi. (8)

In the next section, we shall use these relations to analyze and explain the difficultiesassociated with the solution of discretizations of first-kind Fredholm integral equations.

3.3. SVD analysis and insight

For general matrices, one cannot say much about the singular values and vectors,except that the singular values decay, by definition. However, for matrices that arisefrom the discretization of first-kind Fredholm integral equations, more can be said. Weomit the underlying theory here (see, e.g., [20] for details) and state the most importantresults here.

336 P.C. Hansen / Deconvolution and regularization

1. The singular values of A decay gradually, until they level off at a plateau approxi-mately at the machine precision times σ1 (in infinite precision they would decay tozero).

2. Consequently, the condition number cond(A) = σ1/σn is approximately the recip-rocal of the machine precision which, for practical purposes, can be considered asinfinite.

3. There is no particular gap in the singular value spectrum – typically, the singularvalues follow a harmonic progression σi � i−α or a geometric progression σi �e−αi , where α is a positive real constant.

4. The singular vectors ui and vi have an increasing number of sign changes in theirelements as i increases, i.e., as the corresponding singular values σi decrease. Often,the number of sign changes is precisely i − 1.

To illustrate the above results, we computed the SVD of the matrix A from thegravity surveying problem from section 1, discretized as explained above with n = 64.The singular values are shown in figure 5, and we see that all the singular values indeeddecay gradually (almost geometrically) until they level off at approximately 10−16. Notsurprisingly, the decay rate depends on the depth d – the larger the d, the faster the decay.

The coefficient matrixA in this example is symmetric, and for symmetric matricesthe singular vectors satisfy ui = ±vi . Thus, we only need to consider the left singularvectors ui which are shown in figure 6. We see that the number of sign changes inthe elements of ui is precisely i − 1, supporting our claim that the higher the index i

Figure 5. The singular values of the gravity surveying model problem, discretized with the midpoint quadra-ture rule and n = 64, for two values of the depth d .

P.C. Hansen / Deconvolution and regularization 337

Figure 6. The first nine left singular vectors ui for the gravity surveying problem.

and the smaller the corresponding singular value σi , the more high-frequent the singularvectors ui and vi .

From the above SVD analysis, it is clear that in connection with the analysis andsolution of the problem Ax = bwe should monitor the behavior of the singular values σi ,the SVD-components uT

i b of the right-hand side, and the SVD components uTi b/σi of

the computed “naive” solution. Figure 7 shows these quantities for the gravity modelproblem, and we see that the quantities |uT

i b| decay until they, too, hit the level set bythe machine precision. Note that in the decay region, the quantities |uT

i b| decay strictlyfaster than the singular value, such that in the same region the quantities |uT

i b/σi| alsodecay, only slower. In the region where |uT

i b| levels off, the quantities |uTi b/σi| increase.

The “naive” solution xnaive given by equation (8) is completely dominated by the SVDcomponents corresponding to the smallest singular values, and therefore xnaive appears asa highly oscillatory solution with a large norm ‖xnaive‖2 = 6.1 · 1015. Figure 4 illustratesthis phenomenon.

It is not a coincidence that the absolute values of the SVD components uTi b in this

example decay faster than the singular values. It can be shown that as long as there existsa square integrable solution to the underlying integral equation (i.e., a solution f suchthat

∫ 10 f (t)

2 dt < ∞), then independently of the particular discretization scheme thequantities |uT

i b| will decay faster than the singular values σi (until they eventually hitthe machine precision level). It is beyond the scope of theis paper to go further into theunderlying theory; see [19,20] for details. Before attempting to solve any discretization

338 P.C. Hansen / Deconvolution and regularization

Figure 7. Singular values σi and coefficients |uTib| and |uT

ib|/σi for the gravity surveying model problem.

Apart from rounding errors, there is no noise in the right-hand side b.

(a) (b)

Figure 8. Same plots as before, except that noise with σnoise = 10−6 (a) and σnoise = 10−3 (b) is added tothe right-hand side.

of a first-kind Fredholm equation, one should always try to check whether the quantities|uTi b| indeed decay faster than the singular values – otherwise there is no point in trying

to solve the problem.We now repeat the SVD analysis of the same problem as before, but with noise

added to the right-hand side b = b+e, where the noise vector e has normally distributedelements with zero mean and standard deviation σnoise. Figure 8 shows plots of the SVDquantities σi , uT

i b, and uTi b/σi for two noise levels, σnoise = 10−6 and σnoise = 10−3.

We see that as the noise level increases, the decay region of the quantities |uTi b| gets

narrower, and the norm of the “naive” solution increases.

P.C. Hansen / Deconvolution and regularization 339

What happens in these plots is obviously that for small indices i, the SVD com-ponents uT

i b are dominated by the exact parts uTi b, while for larger indices in the region

where |uTi b| has leveled off, the SVD components uT

i b are dominated by the noise partuTi e. In this last region, the information in the exact right-hand side b is lost, due to the

noise. Moreover, the number of lost components increases as the noise level increases.Consequently, only the first SVD components uT

i b/σi of the “naive” solution carry infor-mation about the desired solution. Hence, all numerical algorithms – direct or iterative –that attempt to compute xnaive fail to compute a reasonable approximation to the desiredsolution; and increasing the problem size n or the machine precision do not provide anyimprovements.

4. Regularization

It is now the time to present various algorithms for stabilizing the computed so-lution, such that it becomes less sensitive to the perturbations. This process is calledregularization, and the corresponding algorithms are called regularization algorithms.

4.1. A simple approach: truncated SVD

It should be clear from the discussion in the previous section that although the“naive” solution xnaive = A−1b is useless, because it is dominated by the contributionsfrom the errors in the right-hand side, some of the SVD components still carry usefulinformation about the desired solution. This information is associated with the first SVDcomponents, corresponding to the largest singular values. The question is how to extractthis information, while discarding the remaining, erroneous SVD components.

Clearly, a simple “brute force” approach to achieve this is to actually compute theSVD and neglect all the undesired SVD components. This type of regularization is calledtruncated SVD (TSVD), and the TSVD solution is computed as

xk =k∑i=1

uTi b

σivi, (9)

where the truncation parameter k must be chosen appropriately, e.g., from a plot of theSVD quantities, such that only the desired SVD components are retained in the TSVDsolution xk . The TSVD method is implemented in REGULARIZATION TOOLS as thefunction tsvd.

In spite of its simplicity, this regularization method has been used successively ina variety of applications. Figure 9 shows some TSVD solutions to the gravity modelproblem with σnoise = 10−3, along with the exact solution. Notice how xk improves as kincreases, until the noise contributions take over for k � 14.

The choice of the truncation parameter k is quite straightforward. Monitor thequantities |uT

i b| and choose k at the transition between the decaying region and the flat re-gion. This ensures that xk consists mainly of those SVD components that can be trusted;

340 P.C. Hansen / Deconvolution and regularization

Figure 9. TSVD solutions xk to the gravity surveying model problem with σnoise = 10−3 for k = 2,4, . . . , 16, and the exact solution.

cf. the discussion in the previous section. There are also other techniques available thatdo not require monitoring of the |uT

i b|-quantities; but this topic lies outside the scope ofthis paper, see instead [20].

We emphasize that the TSVD method is only useful when it is reasonable to com-pute the SVD. Indeed, this is possible for small problems – but as the problem sizeincreases, it becomes prohibitive to compute the SVD.

4.2. Tikhonov regularization

This algorithm was developed independently by Phillips [31] and Tikhonov [33].It is most commonly referred to as Tikhonov regularization, and occasionally as dampedleast squares. The key idea is to accept a nonzero residual Ax − b and in return obtain asmaller solution norm. This problem can be formulated as follows

min{‖Ax − b‖2

2 + λ2‖x‖22

}, (10)

where the regularization parameter λ controls the weight given to minimization of thesolution norm ‖x‖2 relative to minimization of the residual norm ‖Ax − b‖2. It canbe shown that there is always a unique solution xλ to the above problem, which wedenote the Tikhonov solution. Note that as λ approaches zero, the Tikhonov solution xλapproaches the “naive solution” xnaive (or the least squares solution, if A is rectangular),

P.C. Hansen / Deconvolution and regularization 341

while xλ → 0 as λ → ∞. In between, there is (hopefully) a range of λ-values forwhich there is a reasonable balance between the residual and solution norms, and forwhich the Tikhonov solution xλ is a reasonable approximation to the desired solution.In REGULARIZATION TOOLS, Tikhonov regularization is implemented in the functiontikhonov.

To obtain a deeper insight into the properties of the Tikhonov solution, we canexpress xλ in terms of the SVD of A as follows:

xλ =n∑i=1

σ 2i

σ 2i + λ2

uTi b

σivi .

The quantities fi = σ 2i /(σ

2i + λ2) are called the Tikhonov filter factors. They all satisfy

0 � fi � 1, and they control the damping of the individual SVD components of the solu-tion xλ. Specifically, if λ is fixed somewhere between σ1 and σn, then for σi � λwe havefi = 1+ O(λ2/σ 2

i ) � 1, while for σi � λ we have fi = σ 2i /λ

2 + O(σ 4i /λ

4) � σ 2i /λ

2.For singular values σi near λ, the filter factors are in a transition region between the twoabove extremes. Thus, we see that the first SVD components, corresponding to singularvalues greater than λ, contribute with almost full strength to the Tikhonov solution xλ.Similarly, the last SVD components corresponding to singular values smaller than λ aredamped considerably and therefore contribute very little to xλ. Hence, we would expectthat the Tikhonov solutions resemble the TSVD solutions when k and λ are chosen suchthat σk � λ; more details can be found in [18].

Figure 10 shows various Tikhonov solutions to the gravity surveying problem withσnoise = 10−3 for a large range of λ-values – notice the resemblance of these solutionswith the TSVD solutions in figure 9.

Equation (10) is not suited for numerical computations, but there are two otherformulations that lend themselves more to numerical computations. The first of them isthe following least squares formulation

min

∥∥∥∥(AλI)x −

(b

0

)∥∥∥∥2

, (11)

where I is the identity matrix of order n, and the other is the normal equations for thisleast squares problem (

ATA+ λ2I)x = ATb.

While the latter is suited in certain situations, it is the least squares formulation (11) thatis best suited for numerical computations.

When A is symmetric and positive definite, there exists an alternative form ofTikhonov’s method – usually attributed to Franklin [14] – that avoids the normal equa-tions by instead working with the system

(A+ λI)x = b, λ � 0, (12)

342 P.C. Hansen / Deconvolution and regularization

Figure 10. Tikhonov solutions xλ to the gravity surveying model problem with σnoise = 10−3 for λ in therange 10−3–10, plus the exact solution.

which can be solved by means of Cholesky factorization. If A = ∑ni=1wiαiw

Ti is the

eigenvalue decomposition of A, then the regularized solution in Franklin’s method isgiven by x = ∑n

i=1(wTi b)/(αi + λ)wi. This method is particularly useful for Toeplitz

matrices because it preserves the Toeplitz structure of the coefficient matrix.At this stage, we emphasize that the 2-norm of the solution is not always the opti-

mal quantity to include in the Tikhonov formulation. Experience, supported by theory,suggest that for some problems, it is better to minimize the 2-norm of a quantity that ap-proximates a derivative of the underlying solution. For example, if we use the midpointquadrature rule for discretization, then approximations to the first and second derivativesare given by L1x and L2x, respectively, where the two matrices L1 and L2 are given by

L1 = n−1

−1 1. . .

. . .

−1 1

, L2 = n−2

1 −2 1. . .

. . .. . .

1 −2 1

. (13)

Notice that these matrices are not square: L1 is (n − 1) × n and L2 is (n − 2) × n.This leads to the following formulation of Tikhonov regularization in general form, asopposed to the standard form in (10):

min{‖Ax − b‖2

2 + λ2‖Lx‖22

}. (14)

P.C. Hansen / Deconvolution and regularization 343

This general-form version of Tikhonov regularization is also available via the functiontikhonov in REGULARIZATION TOOLS.

4.3. Iterative methods

Iterative regularization methods are important for treating large-scale problems,for which the direct computations involved in, e.g., the SVD algorithm and in Eldén’salgorithm, become prohibitive. Iterative methods only “touch” the coefficient matrix Avia matrix–vector multiplication, and these methods are therefore well suited for sparseand structured matrices.

We start with a classical stationary method called Landweber iteration (developedindependently also by Cimino, Fridman, Picard, Richardson, and many others). Let x(0)

denote the starting vector, and often x(0) = 0; then Landweber iteration takes the form

x(k) = x(k−1) + ωAT(b − Ax(k−1)

), k = 1, 2, 3, . . . ,

where ω is a real positive parameter satisfying 0 < ω < 2/‖ATA‖2, whose value con-trols the behavior of the iterations. If we insert the SVD of A into the above formula,then we can show that the kth iteration vector is given by

x(k) =n∑i=1

(1− (1− ωσ 2

i

)k)uTi b

σivi,

showing that the filter factors for this method are given by f (k)i = 1 − (1 − ωσ 2i )k, for

i = 1, . . . , n. Plots of these filter factors are shown in figure 11 for k = 10, 20, 40and 80. We see that the filter factors resemble those for Tikhonov regularization, with

Figure 11. The function 1 − (1 − σ 2)k which defines the Landweber filter factors, for ω = 1 and k = 10,20, 40 and 80.

344 P.C. Hansen / Deconvolution and regularization

f(k)i � 1 for σ 2

i � 1/ω and f (k)i � kωσ 2i for σ 2

i � 1/ω. This analysis shows thatthe iteration number k plays the role of the regularization parameter: initially, only thelargest SVD components are included in the iteration vector x(k), and as we increase thenumber of iterations we include smaller and smaller SVD components into the iterationvector. The analysis also reveals that Landweber’s method converges slowly: doublingthe number of iterations merely halves the size of the small filter factors.

Similar stationary iterative methods can be found in the literature, some of themextensions of the classical Landweber method, and all of them simple to implement andanalyze (because the iteration matrix is independent of the right-hand side). Unfortu-nately, all them share essentially the same slow convergence.

Instead, we turn our attention to the use of the conjugate gradient (CG) algorithm.In connection with regularization problems, we need a variant of CG that solves thenormal equations ATAx = ATb associated with a least squares problem min ‖Ax− b‖2.This variant is called CGLS, and it was described in the original paper by Hestenes andStiefel [22] where the CG method was first published. Again, let x(0) denote the initialguess, and define the two auxiliary vectors r(0) = b−Ax(0) and d(0) = ATr(0). Then theCGLS iterations take the following form for k = 1, 2, . . .

αk = ‖ATr(k−1)‖2

2

‖Ad(k−1)‖22

,

x(k) = x(k−1) + αkd(k−1),

r(k)= r(k−1) − αkAd(k−1),

βk = ‖ATr(k)‖22

‖ATr(k−1)‖22

,

d(k)=ATr(k) + βkd(k−1).

The vector r(k) is the residual vector for the least squares problem, i.e., r(k) = b−Ax(k),while the intermediate vector AT r(k) is the residual vector for the normal equations be-cause AT r(k) = ATb − ATAx(k).

The CGLS algorithm can be used for regularization in two fashions. One is toapply it to the Tikhonov problem in the least squares formulation (11) with A and breplaced by

(A

λI

)and

(b

0

), and in this fashion compute regularized solutions. The other

is to apply CGLS directly to Ax = b, and use the fact that the iteration number k oftenplays the same role as a regularization parameter as in Landweber’s method.

The first approach needs a good preconditioner (see [15, section 10.3]), but noneof the preconditioners for general problems are suited for Tikhonov regularization prob-lems, due to the spectral properties of the coefficient matrix ATA + λ2I . Developmentof specialized preconditioners for this matrix is a topic of current research, and a fewalgorithms for general unstructured matrices have appeared recently [17]. We return topreconditioners for structured matrices in section 5.3.

The second approach is much simpler to use, because the preconditioner is avoided– we merely apply the above CGLS scheme to A and b. Unfortunately, this method is

P.C. Hansen / Deconvolution and regularization 345

Figure 12. CGLS filter factors f (k)i

as functions of the singular values σi for the gravity surveying test

problem with σnoise = 10−3.

hard to analyze theoretically, because it is a nonstationary method (the iteration matrixdepends on the iteration number k) and because the iteration matrix depends on the right-hand side. Once again, we express the iteration vector in terms of the SVD of A, and weobtain

x(k) =n∑i=1

f(k)i

uTi b

σivi,

where the CGLS filter factors f (k)i depend in a nonlinear way on both b and all thesingular values σ1, . . . , σn. It can be show that f (k)i � 1 for the large singular valuesand f (k)i = O(σ 2

i ) for the small singular values (see [20, section 6.3.2]). In the transitionrange, certain filter factors can become slightly larger than one.

Plots of typical filter factors are shown in figure 12 – notice the resemblance withthe Tikhonov filter factors. Also note that for certain values of k, there is a filter factorslightly larger than one. We see once again that the filter factors dampen the SVD com-ponents corresponding to small singular values, and the transition value of σi at whichthe damping sets in decreases as k increases. But the converge of the CGLS method isfaster than that of Landweber and other classical stationary methods. One can think ofCGLS as an iterative scheme for computing approximate TSVD or Tikhonov solutions,in which a number 5(k) � k of SVD components are captured in the kth iteration, cf.table 1.

346 P.C. Hansen / Deconvolution and regularization

Table 1Corresponding values of k and 5(k) in the CGLS algorithm for

the problem in figure 12.

k 2 4 6 8 10 12 14 16 18

5(k) 2 4 6 7 8 9 10 11 11

One of the difficulties with this use of CGLS as an iterative regularization methodis to determine under which conditions the algorithm has an intrinsic regularization prop-erty. I.e., for which class of problems do the filter factors look as in figure 12, with a flatregion where f (k)i � 1 and a decaying region where f (k)i = O(σ 2

i ). A full analysis ofthis problem has not been developed yet – partial results can be found in [20], where itis concluded that the desired behavior occurs under the following conditions.

• The singular values should decay gradually to zero.

• The decay should not be too gentle.

• The quantities |uTi b|, i.e., the SVD coefficients of the right-hand side, should decay

faster than the singular values.

All three conditions are usually satisfied in connection with discretizations of ill-posed problems – but on the other hand there is no perfect guarantee to obtain the desiredfeatures of the filter factors. Things are complicated even further in the presence offinite-precision arithmetic.

An analysis of the MINRES algorithm (an iterative algorithm related to CGLS) ispresented by Kilmer and Stewart in [27], where it is demonstrated that for discretizationsof ill-posed problems, this algorithm will capture all the SVD components associatedwith the large singular values before it starts to include the remaining SVD components.It is also shown that the number of captured SVD components in the kth step of thismethod is often much larger than k.

5. Deconvolution in one dimension

We now return to the main theme of this paper, namely, deconvolution problems,and we show how the above general algorithms specialize – in various ways – for thisparticular class of regularization problems. The key observation here is that for convo-lution problems, where the kernel satisfies

K(s, t) = h(s − t),the corresponding matrix A derived from a quadrature rule discretization, with elementsaij = wjh(si − tj ), can be written in the form

A = HW, (15)

P.C. Hansen / Deconvolution and regularization 347

whereW = diag(w1, . . . , wn) is a diagonal matrix consisting of the quadrature weights,and the elements of the matrix H are “samples” of h, i.e.,

hij = h(si − tj ), i, j = 1, . . . , n. (16)

At this stage we will assume that the quadrature points si and the collocation points tjare identical and equidistantly spaced, i.e., si = ti = α + βi, i = 1, . . . , n. In this casethe elements of the matrix H satisfy

hij = h(si − tj ) = h(si+5 − tj+5) = hi+5,j+5for all relevant i, j and 5. This special structure of the coefficient matrix can be used toderive very efficient algorithms. We shall first explore the structure of H , and then turnto the regularization algorithms.

5.1. Toeplitz matrices and their SVD

A Toeplitz matrix T is a matrix whose elements depend only on the difference i−jbetween the indices, i.e., T can be written as

T =

t0 t−1 t−2 . . . t1−nt1 t0 t−1 . . . t2−nt2 t1 t0 . . . t3−n...

......

. . ....

tn−1 tn−2 tn−3 . . . t0

. (17)

Notice that an n × n Toeplitz matrix is characterized by only 2n − 1 different elements(or merely n elements if T is symmetric), and it is this feature that makes it possibleto derive efficient algorithms for Toeplitz matrices. Obviously, the matrix H in equa-tion (16) is a Toeplitz matrix when the quadrature and collocation points are identicaland equidistantly spaced.

Toeplitz matrices T are persymmetric, i.e., they are symmetric across the antidiag-onal. Hence, in addition to the relation tij = ti−5,j−5 = ti−j for all relevant i, j , and 5,their elements satisfy the relation tij = tn−j+1,n−i+1. This can also be expressed as thefact that the Toeplitz matrix with its columns in reverse order is symmetric. Let J denotethe exchange matrix

J = 1

···1

which, when multiplied from the right to T , reverses the order of the columns of T .Thus we have T J = (T J )T = JT T and hence the persymmetry of T can be expressedas T = JT TJ . One of the implications of this property is that the inverse of T isalso persymmetric; to see this note that J 2 = I ⇔ J−1 = J and therefore T −1 =(JT TJ )−1 = J−1(T −1)TJ−1 = J (T −1)TJ .

348 P.C. Hansen / Deconvolution and regularization

The persymmetry of T can be used to derive certain symmetry relations betweenthe singular vectors of T . From the symmetry of the matrix TJ , it follows from (6) that

T J =n∑j=1

wjγjλj (γjwj )T ⇐⇒ T =

n∑j=1

wjγjλj (γjJwj)T,

where the quantities γj = ±1 are chosen to make γjλj positive. The rightmost equationabove is identical to the SVD of T , and we recognize uj = wj and vj = γjJwj =γjJuj . Hence, the elements of the left and right singular vectors are related by

vij = γjun−i+1,j , i, j = 1, . . . , n.

I.e., except perhaps for a sign change, the vector vj is identical to uj with its elementsin reverse order.

If T is also symmetric such that uj = γj vj , where the quantities γj = ±1 aregenerally different from the γj , then additional symmetries occur because uij = γj vij =γj (γjun−i+1,j ) = γj γjun−i+1,j , and similarly for the right singular vectors. We cansummarize these symmetry relations as follows:

|uij | = |un−i+1,j | = |vij | = |vn−i+1,j |, i, j = 1, . . . , n.

In other words, the left and right singular vectors of a symmetric Toeplitz matrix areidentical except perhaps for a sign change, and the sequence of elements in each vectoris symmetric around the middle elements except perhaps for a sign change. We illustratethis by means of a small example:

T =0 1 2

1 0 12 1 0

,

U =0.628 0.707 −0.325

0.460 0 0.8880.628 −0.707 −0.325

, V =0.628 −0.707 0.325

0.460 0 −0.8880.628 0.707 0.325

.Here, u1 = Ju1 = v1 = Jv1, u2 = −Ju2 = −v2 = Jv2, and u3 = Ju3 = −v3 =−Jv3. The singular values of T are σ1 = 2.73, σ2 = 2 and σ3 = 0.732.

5.2. Circulant matrices and convolution

Circulant matrices form a special class of Toeplitz matrices in which the bottomelement of a column “wraps around” as the column is repeated to the right in a down-shifted version. Hence, an n× n circulant matrix C has the form

P.C. Hansen / Deconvolution and regularization 349

C =

c0 cn−1 cn−2 . . . c1

c1 c0 cn−1 . . . c2

c2 c1 c0 . . . c3...

......

. . ....

cn−1 cn−2 cn−3 . . . c0

,i.e., C has n different elements satisfying the relation cij = c(i−j)mod n. Among otherthings, circulant matrices can be used to express the convolution of two periodic signalsin matrix form. Recall that if f and h are two n-periodic signals, then the convolution off and h is a new n-periodic signal g with elements given by gi =∑n−1

j=0 fjh(i−j)mod n.Thus, if h is the first column of Ch, then g is obviously given by

g = Chf. (18)

Note that we can interchange the roles of f and h without changing g, hence there isnothing special about the circulant matrix being made up of h.

An amazing fact about circulant matrices is that all circulant matrices of order nhave the same eigenvectors, and the eigenvalues are given by the DFT of the first columnof C. In order to derive this result, we use the fact that the DFT of an n-vector f can bewritten as f = DFT(f ) = Fnf , where the elements of the complex symmetric matrix Fnare given by

(Fn)ij =(

exp

(−2πı

n

))(i−1)(j−1)

,

where ı is the imaginary unit. The inverse of Fn is given by F−1n = n−1conj(Fn) – notice

the complex conjugation – showing that the IDFT can be written as f = IDFT(f ) =F−1n f = n−1conj(Fn)f . We now insert these two relations into expression (1) for g to

obtain

g= IDFT(DFT(h)� DFT(f )

) = F−1n

(Fnh)� (Fnf )

)=F−1

n

(diag(Fnh)Fnf

) = F−1n diag(Fnh)Fnf.

Via the definition of Fn and the expression

F−1n diag(Fnh)Fn = n−1conj(Fn)diag(Fnh)Fn

it is easy to show that this matrix is circulant. Hence, the expression for g is identicalto the expression in (18) if we choose Ch = F−1

n diag(Fnh)Fn, and since there are onlyn free elements in Ch as well as in F−1

n diag(Fnh)Fn this relation is unique. This showsthat diag(Fnh) is a similarity transform of Ch, which preserves eigenvalues, and hencethe elements of the vector Fnh must be the eigenvalues of Ch. Moreover, the columns ofthe matrix n−1/2Fn are orthonormal eigenvectors of Ch.

At this stage we emphasize that all DFTs and IDFTs should be computed by meansof the FFT algorithm. If n is a power of two, then a complex FFT requires 5n log2 n

350 P.C. Hansen / Deconvolution and regularization

flops, and if data are real then the FFT can be implemented in such a way that only halfas many flops are required. These numbers actually ignore the overhead in computingthe complex exponentials; in Matlab the total amount of flops is 2.5n log2 n+5n+O(1)for real data. Notice that the interchanges of the elements that is usually part of an FFTalgorithm is not necessary in the above application and can therefore be switched off,thus saving computational overhead (although this is not possible in Matlab).

5.3. Matrix–vector multiplication by FFT

In connection with iterative algorithms, it is important to know that matrix–vectormultiplication with a Toeplitz matrix can be performed in O(n log2 n) flops – as opposedto 2n2 flops for a general matrix–vector multiplication. We derive the algorithm here,and more details can be found in [34].

The key idea is to embed the n × n Toeplitz matrix T in a larger p × p circulantmatrix C, and use the FFT algorithm to implement the fast matrix–vector multiplica-tion with C. Specifically, using the notation of equation (17), the first column of C isconstructed such that

C(: , 1) = (t0, t1, . . . , tn−1, 0, . . . , 0, t1−n, . . . , t−1)T.

The middle part of this vector consists of p − 2n + 1 zeros, and p should be chosento be the smallest possible power of 2 satisfying p � 2n in order to speed up the FFTcomputations. For example, if T is the nonsymmetric 3× 3 matrix

T =3 −2 −1

2 3 −21 2 3

then p = 8 and C takes the form

C =

3 −2 −1 0 0 0 1 22 3 −2 −1 0 0 0 11 2 3 −2 −1 0 0 00 1 2 3 −2 −1 0 00 0 1 2 3 −2 −1 00 0 0 1 2 3 −2 −1−1 0 0 0 1 2 3 −2−2 −1 0 0 0 1 2 3

,

where we recognize T as the leading 3× 3 principal submatrix of C.In order to multiply a vector x with T we first pad this vector with p − n zeros to

obtain

x =(x

0

),

P.C. Hansen / Deconvolution and regularization 351

and when we multiply x with C, we obtain the new vector

z = Cx =(T x

z2

).

Due to the zeros in x, the first n components of z are identical to the desired matrix–vector product T x. If data are real and the DFT of C(: , 1) has been precomputed,then the matrix–vector multiplication Cx can be computed very efficiently by means ofessentially two FFTs (one real and one complex) in only about 7.5p log2 p flops. Sincep is bounded above by 4n we obtain the approximate upper bound 30n log2 n for theflop count involved in one matrix–vector multiplication, and if n is a power of 2 (suchthat p = 2n) then the approximate bound reduces to 15n log2 n flops.

Note that if T is real and symmetric, then the vector C(: , 1) is “real even” andDFT(C(: , 1)) is a real vector – but the complexity of the Toeplitz-based matrix–vectormultiplication algorithms stays the same.

Whether the FFT-based matrix–vector multiplication actually pays off depends onthe elements of the Toeplitz matrix T . If T is a banded matrix with bandwidth nb (or ifthe elements outside this band are so small that they can be considered as zeros), thenthe matrix–vector multiplication requires 2nbn flops. When n is a power of two, thebreak-even value of n, for which the two approaches require about the same work, isn = 2nb/7.5 – i.e., for a fixed bandwidth nb the ordinary banded matrix–vector multiplyis faster than the FFT-based approach for n > 2nb/7.5. Figure 13 illustrates this point.

Figure 13. Flop counts for the FFT-based Toeplitz matrix–vector multiplication algorithm, compared to flopcounts for ordinary matrix–vector multiplication with full and banded matrices.

352 P.C. Hansen / Deconvolution and regularization

5.4. Circulant preconditioners

We will here briefly summarize the main ideas of some current algorithms for com-puting circulant preconditioners specialized to regularization problems with a Toeplitzcoefficient matrix. Circulant matrices are a natural choice as preconditioners for Toeplitzmatrices because they are inexpensive to construct and use.

If T is square then we can start by computing a circulant approximation CT to T .For example, we can use the closest circulant matrix to T in the Frobenius norm [9],whose first column has elements given by

(CT )i1 = (n− i)ti + iti−nn

, i = 0, . . . , n− 1. (19)

Next we decompose the circulant matrix CT as

CT = F−1n /T Fn,

where Fn is the Fourier matrix and /T is a diagonal matrix with the eigenvaluesof CT . We also need the following expression for the transpose: CT

T = conj(CTT ) =

conj(Fn/T F−1n ) = conj(Fn) conj(/T )n−1Fn = F−1

n conj(/T )Fn, where we have usedthat F−1

n = n−1conj(Fn). Then it is natural to precondition the Tikhonov problem (10)with the following preconditioner

M = (CTT CT + λ2I

)−1 = (F−1n conj(/T )/T Fn + λ2I

)−1

=F−1n

(|/T |2 + λ2I)−1Fn.

By its definition M is real and symmetric, and if we choose Q as the square root of M,

Q = M1/2 = F−1n

(|/T |2 + λ2I)−1/2

Fn, (20)

then it follows that Q is also real and symmetric, and the preconditioned problem be-comes Q(T TT + λ2I )Qz = QT Tb with x = Qz. This symmetric problem is solvedby means of the CGLS algorithm. We emphasize that the matrix Q is never formedexplicitly, instead FFTs are used to perform all the matrix multiplications. Further im-plementation details and extensions to 2-D problems are described in [8], where thegeneral-form regularization problem in (14) is also considered.

The same idea is used in [16] to derive a regularizing circulant preconditioner forsymmetric matrices. Again we start with a circulant approximation CT = F−1

n /T Fnto T , but now we use the preconditioner

Qτ = F−1n DFn (21)

where D = diag(d1, . . . , dn) is a diagonal matrix with elements

di ={|λi|−1/2 for |λi| � τ,

1 else(22)

in which λi are the eigenvalues of CT . The preconditioned symmetric problemQτTQτz = Qτb with x = Qτz can then be solved by means of CG if T is positive

P.C. Hansen / Deconvolution and regularization 353

definite, or ORTHOMIN if T is indefinite. The threshold τ plays the role of the regular-ization parameter, in the sense that eigenvalues of CT whose magnitude is less than τ arenot affected by the preconditioner, while the remaining eigenvalues of CT are mappedto ±1. We refer to [16] for more details and extensions to the 2-D case.

Finally we mention an algorithm from [28] for a banded (n+q)×n Toeplitz matrixof the form

T =

t0... t0

tq...

. . .

tq. . .

.

The key idea here is to augment T with additional q columns – in the obvious way – tomake it square and circulant,

Caug =(T , T

) = F−1n /augFn.

Following the previously mentioned techniques, we now formally introduce the realsymmetric matrix

Qaug = F−1n

(|/aug|2 + λ2I)−1/2

Fn.

Finally, we partition Qaug as

Qaug =(Q, Q

)such that Q has the same dimensions as T , and solve the preconditioned problemQ(T TT + λ2I )QTz = QT Tb with x = QTz. Again we stress that Qaug or Q arenever formed explicitly, and we refer to the original paper for details and extensions to2-D problems.

We mention in passing that instead of using a circulant preconditioner, one cantransform the Toeplitz matrix into a Cauchy-like matrix (which is numerically stable)and then construct a preconditioner from a rank-revealing pivoted LU decomposition ofthe Cauchy-like matrix; see [26] for details.

5.5. Direct algorithms for Toeplitz matrices

We have shown in the previous section that iterative regularization algorithms canbe implemented very efficiently for convolution problems with Toeplitz matrices. As weshall see in this section, the same is true for direct methods implementing Tikhonov’smethod. Since the diagonal weight matrix W in (15) can always be “absorbed” into thesolution vector x, we can assume without loss of generality that the coefficient matrix Tis a Toeplitz matrix.

We consider first the special case where the coefficient matrix T is triangular andToeplitz, and L is a p × n Toeplitz matrix with dimensions p � n and with zeros below

354 P.C. Hansen / Deconvolution and regularization

the main diagonal. Triangular Toeplitz matrices arise, e.g., in connection with Volterra-type deconvolution problems (4), and L is almost always a banded Toeplitz matrix whenthe integral equation is discretized with equidistant abscissas. We shall here consider thecase where T is upper triangular; problems with a lower triangular coefficient matrix caneasily be brought into this form. The coefficient matrix in the least squares formulationof Tikhonov regularization then takes the following from (shown here for n = 6 andp = 4, and ignoring the possible band structure of L):

(T

λL

)=

t0 t1 t2 t3 t4 t5t0 t1 t2 t3 t4

t0 t1 t2 t3t0 t1 t2

t0 t1t0

l0 l1 l2 l3 l4 l5l0 l1 l2 l3 l4

l0 l1 l2 l3l0 l1 l2

We shall now demonstrate how to transform this matrix into upper triangular form

by means of n Givens transformations. The algorithm is due to Eldén [11]. First, weapply a Givens rotation to rows 1 and n + 1 in order to annihilate the element l0; thisrotation changes all the elements in the two rows, and we denote the new elements witha prime, e.g., the (1, 1)-element becomes a′0. The same rotation is applied (implicitly) tothe row pairs (j, n + j) for j = 2, . . . , p, and due to the Toeplitz structure of T and Lthe new rows are just shifted versions of rows 1 and n + 1. Thus, we obtain the firstintermediate matrix

M ′ =

t ′0 t ′1 t ′2 t ′3 t ′4 t ′5t ′0 t ′1 t ′2 t ′3 t ′4

t ′0 t ′1 t ′2 t ′3t ′0 t ′1 t ′2

t ′0 t ′1t ′0

l′1 l′2 l′3 l′4 l′5l′1 l′2 l′3 l′4

l′1 l′2 l′3l′1 l′2

.

We now apply the same procedure to the submatrix M ′(2:n+ p, 2:n) to obtain the sec-ond intermediate matrix M ′′, where a double prime denotes an element that is changedin the second step:

P.C. Hansen / Deconvolution and regularization 355

M ′′ =

t ′0 t ′1 t ′2 t ′3 t ′4 t ′5t ′′0 t ′′1 t ′′2 t ′′3 t ′′4

t ′′0 t ′′1 t ′′2 t ′′3t ′′0 t ′′1 t ′′2

t ′′0 t ′′1t ′′0

l′′2 l′′3 l′′4 l′′5l′′2 l′′3 l′′4

l′′2 l′′3l′′2

.

Obviously, we can repeat this process on the submatrix M ′′(3:n + p, 3:n), and so on,until we end up with an upper triangular matrix after n steps. The algorithm uses solelyGivens rotations and is therefore numerically stable. The total amount of work is approx-imately 8n2, and the storage requirement is dominated by the need to store n(n + 1)/2elements of the final triangular matrix.

Next we consider the more general case where T is a generalm×n Toeplitz matrix,and we present an algorithm due to Bojanczyk et al. [6] (with improvements by Parkand Eldén [29]) for computing the triangular matrix R in the QR factorization(

T

λI

)= QR.

The algorithm can be augmented to treat general L matrices. To get started, we partitionthe coefficient matrix in the following ways

(T

λI

)=

t0 uT

v T

λ 0T

0 λI

=T u

v T tm−nλI 00T λ

,where T is the leading/trailing (m−1)× (n−1) submatrix of T , I is the identity matrixof order n− 1, and the four vectors u, v, u, and v are chosen to fill out T correctly. Wealso partition the n×n upper triangular matrix R accordingly in the following two ways

R =(r11 zT

0 Rb

)=(Rt z

0T rnn

).

By comparing the submatrices in the equation(T

λI

)T (T

λI

)= RTR

we are lead to the following relations for the three nonzero blocks of R:

r211 = t20 + vTv + λ2, zT = t0u

T + vT T

r11, RT

b Rb = RTt Rt + uuT − v vT − zzT.

356 P.C. Hansen / Deconvolution and regularization

Notice that λ enters only explicitly in the formula for r11.We can use the first two relations to immediately compute the first row of R,

namely, (r11, zT). The third relation shows that the submatrix Rb is related to Rt via

a sequence of three rank-one modifications. Hence, if we knew Rt then we could com-pute Rb by means of well-established numerical techniques for up- and downdating, see,e.g., [15, section 12.5].

The heart of the algorithm is to recognize that we already know the first row of Rt ,which is identical to R(1, 1:n − 1), and with this information we can compute the firstrow of Rb. First, we determined a Givens rotation G1 which, when applied to the firstrow of Rt and uT, annihilates the first element of u, i.e.,

G1

(Rt

uT

)=(Rt(u′)T

), u′(1) = 0.

Note that Rt is identical to Rt except for its first row. Next, we determine a secondGivens rotation G1 which, when applied to vT and zT, annihilates the first element of z,i.e.,

G1

(vT

zT

)=(vT(z′)T

), z′(1) = 0.

Finally, we determine a stabilized hyperbolic rotation H1 which, when applied to thefirst row of Rt and vT, annihilates the first element of v, i.e.,

H1

(Rt

vT

)=(R′t(v ′)T

), v ′(1) = 0.

We have now finished the computation of the first row of Rb and thus the second row ofthe desired triangular matrix R.

To proceed, we continue the up- and downdating process on the second row of R′t .First, we use two Givens rotations G2 and G2 to annihilate the second element of u′ bymeans of the second row of R′t , and to annihilate the second element of z′ by means ofv′, respectively:

G2

(R′t(u′)T

)=(R′t(u′′)T

), G2

((v′)T(z′)T

)=((v′)T(

z′′)T

),

where u′′(1) = z′′(1) = 0. Then we use a stabilized hyperbolic rotation H2 to annihilatethe second element of v′ by means of the second row of R′′t :

H2

(R′t(v′)T

)=(R′′t(v ′′)T

), v′′(1) = 0.

At this stage, we have computed the first three rows of the desired matrix R, and it isclear that this process can be repeated until all of R has been computed. The algorithm

P.C. Hansen / Deconvolution and regularization 357

requires mn + 6n2 flops. Information about Q is not explicitly available, so we mustcompute the Tikhonov solution via the semi-normal equations (see [5, section 6.6.5]):

RTRxλ = T Tb.

This requires additional (m + 2n)n flops, which can be reduced if the FFT is used tocompute T Tb. For practical issues as well as an error analysis, see [30].

A word about the numerical stability of the algorithm. It is proved in [29] that theforward error in the computed R is proportional to the condition number of the matrixT TT +λ2I . Although T is very ill conditioned, λ is always chosen such that T TT +λ2I

is more well conditioned, and hence R is normally computed with sufficient accuracy.The paper [29] also describes an algorithm to improve the accuracy of R.

If T is symmetric and positive definite then as an alternative we can use Franklin’smethod (12) and solve the system (T + λI)x = b for positive λ. This system can besolved in 6n2 flops by means of Levinson’s algorithm, cf. [15, section 4.7]. Franklin’smethod may also be applied to symmetric indefinite or general problems, but then caremust be taken to ensure that λ is sufficiently different from all eigenvalues of T , and itmay be necessary to use a look-ahead Levinson algorithm [10].

We emphasize that all algorithms in this section must be started over if a regularizedsolution with a new value of λ is desired.

5.6. Periodic versus nonperiodic deconvolution problems

The periodic deconvolution problem (1) corresponds to a linear system of equationswith a circulant coefficient matrix C, and if regularization is not needed then the solutionis given by equation (2). If regularization is needed, then we can easily modify thisformula to include Franklin’s version of regularization (12). Since C = W/WT withW = n−1/2Fn and / = diag(DFT(h)), it is easy to see that the regularized solution canbe computed as

f = IDFT(DFT(g)� (DFT(h)+ λ)). (23)

Although Franklin’s method is not guaranteed to work for general matrices, it will workfor any circulant C because its eigenvectors (the columns of Fn) have the characteristicoscillation property mentioned in section 3.3.

In nonperiodic deconvolution problems where the Toeplitz coefficient matrix Tis not circulant, i.e., where the signal h is not periodic with period n, equation (23) isoccasionally still used to compute a solution in order to achieve the computational speedof the FFT algorithm, at the cost of a degradation of the computed solution.

Inspired by the FFT-based technique to compute the Toeplitz matrix–vector productfast, a typical approach is to embed T into a larger circulant matrixC as described above,pad the right-hand side b with additional p − n zeros to make it conform with C, andthen compute the vector x = C−1b, where b is the augmented version of b. Then as theapproximate solution x we use those elements of x that correspond to the locations ofthe b-element in b. We note that if a cyclic shift is applied to b then the same cyclic shift

358 P.C. Hansen / Deconvolution and regularization

applies to x, and thus the approximate solution x is independent on the locations of thezeros in b.

Unfortunately, this approach is not guaranteed to yield a good approximation to thedesired solution. To illustrate this, we assume that n is a power of 2 and introduce thenotation

C =(T ϒ

ϒ T

), x =

(x

z

), b =

(b

0

).

Then it is easy to show that the inverse of C is given by

C−1 =(T −1 0

0 0

)+(T −1ϒS−1ϒT −1 −T −1ϒS−1

−S−1ϒT −1 S−1

), (24)

where the Schur complement S is given by S = T −ϒT −1ϒ . Hence, it follows that theapproximation x can be expressed as

x = x + T −1ϒS−1ϒx

showing that there is no guarantee that x resembles the exact solution x.The error component T −1ϒS−1ϒx often appears as artificial oscillations, called

“ringing”, at both ends of the approximate solution x. We illustrate this with a numericalexample with n = 64, a Gaussian kernel chosen such that aij = exp(−(i − j)2/2), anda piecewise constant solution x. The right-hand side b = T x is padded with 32 zeroson top and 32 zeros below, and the computed vector x of length 2n = 128 is shownin figure 14 with the exact solution x superimposed at the correct position. Notice thelarge, undesired ringing.

Figure 14. The approximate solution x is embedded in the vector x, and the exact solution x is superim-posed. Notice the large oscillations (“ringing”).

P.C. Hansen / Deconvolution and regularization 359

Alternatively, if the underlying signal h is peaked at index 0 and decaying awayfrom this index, thus making T as effectively banded matrix, then one could re-place h with a periodic signal of the same length, i.e., with the sequence h0, h1,

. . . , hn/2−1, hn/2, hn/2−1, . . . , h2, h1. This correspond to replacing the Toeplitz matrix Twith a circulant matrix of the same order whose first column – in Matlab notation –is given by [T (1:n/2, 1);T (n/2 + 1:−1: 2)]. The solution to this modified problemhas exactly the same difficulties with ringing as the solution in the previously describedapproach.

The ringing effect is caused by the sharp transitions in the padded vector b betweenthe nonzero elements in b and the padded zeros. Figure 15 illustrates this: the top plotshows b while the solid line in the bottom plot is the correct right-hand side vector givenby

C

(x

0

)=(T x

ϒx

)=(b

ϒx

). (25)

Both plots are circularly shiftet (as in figure 14) in order to emphasize the error, whichis shown as the black filled areas. It is precisely this error, caused by the “missing”

(a)

(b)

Figure 15. (a) The padded right-hand side b, and (b) the correct right-hand side together with their differ-ence. Both figures are circularly shifted to emphasize the error which occurs at the borders of the padded

parts.

360 P.C. Hansen / Deconvolution and regularization

(a) (b)

(c) (d)

Figure 16. Figures (a) and (c) show the coefficients when expressing the correct right-hand side and thecorresponding solution in terms of the eigenvectors of C (i.e., the columns of Fn); (b) and (d) show the

same coefficients for the padded right-hand side b and the corresponding approximate solution x.

subvector ϒx in b, that causes the oscillations in x, because high-frequency componentsare needed in order to represent the error and because these high-frequncy componentsare amplified by the inversion process, thus causing the ringing effect.

Figure 16 illustrates the same phenomenon from a different perspective, namely, interms of the coefficients in an eigenvector expansion. Recall that the eigenvectors of Care the columns of the DFT matrix Fn. Figures 16(a) and (c) show the absolute valuesof the coefficients for the correct right-hand side and the corresponding correct solution,given by the vectors

Fn

(b

ϒx

)and Fn

(x

0

)= /−1Fn

(b

ϒx

).

Note that the correct right-hand side coefficients decay fast enough to ensure that theinversion (i.e., the multiplication with /−1) produces a solution with decaying coeffi-cients.

P.C. Hansen / Deconvolution and regularization 361

Figures 16(b) and (d) show the absolute values of the coefficient vectors Fnb andFnx = /−1Fnx for the padded right-hand side b and the corresponding approximatesolution x. The perturbation of the high-frequent components is clearly visible. Thedecay of the coefficients Fnb is now slower than the decay of the eigenvalues for indicesgreater than 50. As a consequence the coefficients to the high-frequency components ofthe approximate solution x increase and give rise to a severe perturbation of x.

5.7. An iterative solver based on embedding

We shall here describe a new iterative algorithm2 for solving Toeplitz systems,based on the augmented matrix C and using the fact that a system involving this matrixcan be solved in O(n log2 n) operations. Again, we assume that the order n is a power oftwo.

Our outset is the relation (25) which shows that if we could somehow “plug in” thevector ϒx in the bottom half of the augmented right-hand side, then the top half of thesolution vector would consist of the solution x. Now if x0 denotes an initial guess forthe solution (which may be the zero vector), and if we insert ϒx0 in the right-hand side,then it follows from (24) that the corresponding solution is given by

x1 =(x1

z1

)= C−1

(b

ϒx0

)=(x

0

)+(T −1ϒS−1ϒ

(x − x0

)−S−1ϒ

(x − x0

) ).

Recall that S is the Schur complement given by S = T −ϒT −1ϒ .Assuming that x1 is a better approximation to x than x0 we can repeat the process,

i.e., insert ϒx1 in the right-hand side and solve the augmented system, which producesthe next approximation x2 = x + T −1ϒS−1ϒ(x − x1). Repeating this scheme in aniterative fashion – each time solving the augmented system by means of FFT – leads toa sequence of iterates formally given by

xk = x + T −1ϒS−1ϒ(x − xk−1

) = x − (−T −1ϒS−1ϒ)k(x − x0

).

The error in the kth iterate is therefore given by

x − xk = (−M)k(x − x0

), k = 1, 2, . . . ,

where we have introduced the matrix

M = T −1ϒS−1ϒ.

Obviously, a sufficient condition for the iterates to converge to x is that the spectralradius ρ(M), i.e., the largest of the absolute values of the eigenvalues of M, is strictlyless than one.

In order to analyze this condition in detail, we write S as

S = T −ϒT −1ϒ = T (I − T −1ϒT −1ϒ)

2 The algorithm is based on an idea by J. Hald, Brüel and Kjær, Denmark.

362 P.C. Hansen / Deconvolution and regularization

to obtain

S−1 = (I − (T −1ϒ

)2)−1T −1.

Inserting this into the expression for M we get

M = T −1ϒ(I − (T −1ϒ

)2)−1T −1ϒ.

The next step is to introduce the eigenvalue decomposition of T −1ϒ , given by

T −1ϒ = WDW−1, D = diag(δ1, . . . , δn).

Inserting this into the above formula for M leads to

M = WD2(I −D2

)−1W−1

showing that the eigenvalues of M are δ2i (1 − δ2

i )−1 for i = 1, . . . , n. The requirement

ρ(M) < 1 thus leads to the requirement∣∣∣∣ δ2i

1− δ2i

∣∣∣∣ < 1, i = 1, . . . , n.

Figure 17. The first nine iterates from the iterative algorithm based on embedding the Toeplitz matrix intoa larger circulant matrix.

P.C. Hansen / Deconvolution and regularization 363

A straightforward computation shows that if δi = �i + i�i then∣∣∣∣ δ2i

1− δ2i

∣∣∣∣2 = |δi|4|δi|4 + 1− 2(�2

i − �2i )

and for this quantity to be less than one we must require that 1−2(�2i −�2

i ) > 0, leadingto the requirement �2

i − �2i < 1/2.

We have thus proved that the iteration described above will converge to x providedthat all the eigenvalues δi = �i + i�i of the matrix T −1ϒ satisfy the requirement

Real(δ2i

) = �2i − �2

i <1

2, i = 1, . . . , n.

It remains to give an interpretation of the matrix T −1ϒ and its eigenvalue; we have notfound such an interpretation yet.

We illustrate the above iterative algorithm with the gravity surveying problem usingn = 16 and d = 0.1, for which ρ(M) = 0.78. The first nine iterates are shown infigure 17; notice that the middle entries in the solution vector xk converge faster than theentries at the ends.

6. Deconvolution in two dimensions

In this chapter we discuss certain numerical aspects of 2-D convolution problems.Such problems arise, e.g., in image processing, and the dimensions of these problemsquickly get large – for example, 1024× 1024 images are now common in astronomy aswell as other areas.

Although the underlying techniques and difficulties are essentially the same as inthe 1-D case, there are certain new techniques that come in very handy for 2-D problems.We start with a brief introduction to the world of 2-D convolution problems, and thenwe introduce a very useful tool from linear algebra, the Kronecker product. After that,we turn to the discretization and numerical treatment and solution of 2-D problems.

6.1. 2-D deconvolution problems

The general 2-D version of a first-kind Fredholm integral equation takes the form∫ 1

0

∫ 1

0K(x, y, x′, y′

)f(x′, y′

)dx′ dy′ = g(x, y). (26)

We shall limit our discussion here to the important case where the kernel K is a realconvolution operator K(x, y, x′, y′) = h(x − x′, y − y′) whose variables x − x′ andy − y′ separate, i.e., the kernel has the special form

K(x, y, x′, y′

) = κ(x − x′)ω(y − y′),

364 P.C. Hansen / Deconvolution and regularization

where κ and ω are real functions. When we insert this product into the general formof the integral equation, we notice that due to the separation of the variables in K, theintegration can be split into the two variables x′ and y′:∫ 1

0κ(x − x′)(∫ 1

0ω(y − y′)f (x′, y′) dy′

)dx′ = g(x, y). (27)

This Fredholm integral equation shares exactly the same analytical and numerical diffi-culties as the 1-D version.

We give an example of such a problem arising in confocal microscopy [4], a tech-nique that provides improved resolution compared to conventional light microscopy. Ifwe assume that the object is uniformly illuminated, and that the collector lens is sim-ply an aperture of width w, then the 2-D object f is related to its image g via a 2-Dconvolution equation whose kernel separates as in (27), with functions κ and ω given by

κ(z) = ω(z) = sin(πwz)

πwz.

Precisely the same 2-D integral equation arises in connection with extrapolation of band-limited signals [32]. Other 2-D convolution problems with separable kernels arise inconnection with image restoration problems, and we return to this subject shortly.

6.2. Kronecker products

Given two matrices A and A of dimensions m × n and m × n, respectively, theKronecker product A⊗ A is defined as a new matrix of dimensions mm× nn given by

A⊗A =

a11A a12A . . . a1nA

a21A a22A . . . a2nA...

......

am1A am2A . . . amnA

. (28)

The Kronecker product can also be expressed in terms of the columns of A = (a1,

. . . , an) and A = (a1, . . . , an) as

A⊗ A = (a1 ⊗ a1, . . . , a1 ⊗ an, a2 ⊗ a1, . . . , an ⊗ a1, . . . , an ⊗ an). (29)

The Kronecker product is very useful when dealing with discretizations of 2-D problemswith separable kernels, and it enjoys a number of properties that come in handy in thisconnection, such as the relations

(A⊗ B)T = AT ⊗ BT (30)

and

(A⊗ B)(C ⊗D) = (AC)⊗ (BD). (31)

Some norm relations for Kronecker products are

‖A⊗ B‖p = ‖A‖p‖B‖p, p = 1, 2,∞,F.

P.C. Hansen / Deconvolution and regularization 365

More details about Kronecker products can be found in [34].Along with the Kronecker product, we need the “vec” notation. If X is an m × n

matrix with column partitioning X = (x1, . . . , xn), then we define the vector vec(X) oflength mn as

vec(X) =x1...

xn

. (32)

The Kronecker product and the “vec” notation are connected via the following importantrelation, which we shall make extensive use of(

A⊗ A)vec(X) = vec(AXA

T). (33)

One of the many amazing properties of the Kronecker product A ⊗ A is that itssingular value decomposition is entirely made up of the SVDs of the two matrices A andA (and the same is true for the eigenvalue decomposition). Specifically, if both A and Aare square, and if their SVDs are given by

A = U(V T and A = U ( V T,

then it follows from (29) and (31) that

A⊗ A = (U ⊗ U)(( ⊗()(V ⊗ V )T =n∑i=1

n∑j=1

(ui ⊗ uj

)(σ iσj

)(vi ⊗ vj

)T

which, except for the ordering of the singular values and vectors, constitutes the SVD ofA ⊗ A. We see that the nn singular values of A ⊗ A consist of all the products of thesingular values of A and A, while the left and right singular vectors of A⊗A consist ofall the Kronecker products of the left and right singular vectors of A and A. Hence, towork with the SVD of A⊗ A, one merely needs to compute the two SVDs of A and A.As we shall see below, similar computational savings can always be made when workingwith Kronecker products.

6.3. Discretization when variables separate

After the above venture into the world of Kronecker products, we now turn ourattention to the 2-D convolution problem in (27). To discretize this 2-D deconvolutionproblem, we use (for simplicity) the midpoint quadrature rule. First, we apply this ruleto the “inner” integral at the quadrature points y′5, 5 = 1, . . . , n, thus arriving at thefunction φ:∫ 1

0ω(y − y′)f (x′, y′) dy′ � n−1

n∑5=1

ω(y − y′5

)f(x′, y′5

) = φ(x′, y).

366 P.C. Hansen / Deconvolution and regularization

Next, we apply the midpoint quadrature rule to the “outer” integral at the quadraturepoints x′k , k = 1, . . . , n, thus arriving at the function ψ :∫ 1

0κ(x − x′)φ(x′, y) dx′ � n−1

n∑k=1

κ(x − x′k

)φ(x′k, y

) = ψ(x, y).Then we use collocation to obtain a system of linear equations, and we use as manycollocation points as quadrature points to make the system square,

ψ(xi, yj ) = g(xi, yj ), i, j = 1, . . . , n.

Notice that we have used the same number of quadrature and collocation points in eachvariable – this is not required, but makes our exposition simpler.

We shall now derive the equations for the system of linear equations correspondingto the above discretization scheme. First, we need to introduce the four n × n matricesA, A, F , and G with elements given by

Aik = n−1κ(xi − x′k

), Aj5 = n−1ω

(yj − y′5

)Fk5= f

(x′k, y

′5

), Gij = g(xi, yj ),

where all indices are in the range 1, . . . , n. Note that A and A consist of samples of thefunctions κ and ω, respectively, while F and G consist of samples of the approximatesolution f and the right-hand side g, respectively. We can now define an n×nmatrixJthat corresponds to the “inner” integration, with elements given by

Jkj = φ(x′k, yj

) = n−1n∑5=1

ω(yj − y′5

)f(x′k, y

′5

), j, k = 1, . . . , n,

and by studying the indices of the above expression it follows that J can be written asJ = FAT

, where F andA are defined above. Similarly, we can define an n×nmatrixKthat corresponds to the “outer” integration, with elements given by

Kij = ψ(xi, yj ) = n−1n∑k=1

κ(xi − x′k

)φ(x′k, yj

), i, j = 1, . . . , n,

and this matrix can be written as K = AJ, where A is also defined above. Collocationthen corresponds to the requirement that K = G, and in this way we have arrived at thefollowing linear relations between the solution and the right-hand side:

AFAT = G ⇐⇒ (

A⊗ A)vec(F ) = vec(G). (34)

The rightmost equation is a system of n2 linear equations in n2 unknowns, and corre-sponds to the linear system Ax = b in the 1-D case. This system is useful for theoreticalconsiderations, while the leftmost expression in (34) is more suited for numerical com-putations. We note that if both A and A are Toeplitz matrices then the coefficient matrixA⊗ A is a block Toeplitz matrix with Toeplitz blocks.

P.C. Hansen / Deconvolution and regularization 367

6.4. Digital image reconstruction

Digital image restoration – in which a noisy, blurred image is restored on the basisof a mathematical model of the blurring process3 – is probably the most well-knownexample of a 2-D deconvolution problem. A recent survey of this topic, including adiscussion of many practical aspects, can be found in [3]; here we limit our discussion tothose aspects of image restoration that are directly connected with numerical inversionalgorithms.

A digital gray-scale image is really an m× n matrix F whose nonnegative entriesrepresent light intensity, and today’s digital images are usually recorded by means of aCCD camera that records the number of photons hitting a 2-D array of sensors. Hence,we can consider the image F as a sampled version of a continuous 2-D function f (x′, y′)representing the exact image, and this models fits with the above-mentioned discretiza-tion scheme based on the midpoint quadrature rule. Digital images are usually storedas integers using 8 or 10 bits, but in connection with our deconvolution algorithms weconsider the matrix elements as real numbers.

The noise in a digital image typically consists of photon noise (usually with a Pois-son distribution) arising from the discrete nature of the light, plus amplifier noise (usuallycolored Gaussian noise) from the electronic components that convert the recorded lightintensity into a digital signal.

There are many sources of blur, and some of the most important are discussed herealong with their point spread functions, i.e., the kernels K that model the blurring of theunderlying exact image f . We focus on spatially invariant point spread functions whosekernels are separable and of convolution type, i.e., K(x, y, x′, y′) = κ(x−x′)ω(y−y′).This means that the blurring is identical in all parts of the image, and separates into purehorizontal and pure vertical components.

Motion blur arises when the camera moves while the image is recorded, or whenthe object itself moves while the camera is steady. As long as the exposure time is small,we can consider the motion of the camera or the object as linear, and if the motion ishorizontal then the corresponding point spread function takes the form

K(x, y, x′, y′

) = hL(x − x′) = {(2L)−1,∣∣x − x′∣∣ � L,

0, else,(35)

where the parameter L characterizes the smearing of the image. Similarly, if the motionis vertical then the point spread function is given by

K(x, y, x′, y′

) = hL(y − y′) = {(2L)−1, |y − y′| � L,

0, else,

where L plays the same role as before. The corresponding discretizations A ⊗ A ofthese two point spread functions are given by B5 ⊗ I and I ⊗ B5, respectively, where I

3 Opposed to image enhancement where the image is manipulated on a heuristic basis.

368 P.C. Hansen / Deconvolution and regularization

is the identity matrix and B5 is a banded Toeplitz matrix of appropriate dimensions withelements given by

(B5)ij ={(25)−1, |i − j | � 5,

0, else,(36)

where the parameter 5 depends on both L and the discretization.Out-of-focus blur arises, of course, when the lens is out of focus, i.e., when the

focal point of the lens does not match with the light-sensitive CCD. A good model ofthe point spread function for out-of-focus blur is

K(x, y, x′, y′

) = {(πR2)−1,

√(x − x′)2 + (y − y′)2 � R,

0, else,(37)

where the parameter R characterizes the defocus. This function, however, does notseparate in the variables x − x′ and y − y′, and it is therefore customary to use analternative point spread function whose variables do separate, namely,

K(x, y, x′, y′

) = hR(x − x′)hR(y − y′),where the function hR is identical to the horizontal/vertical motion blur function (35)with L replaced by R. The corresponding discretization of the latter point spread func-tion is given by A⊗ A = B5 ⊗ B5.

Atmospheric turbulence blur arises, e.g., in remote sensing and astronomical imag-ing, and is due to long-term exposure through the atmosphere where turbulence in theatmosphere gives rise to random variations in the refractive index. For many practicalpurposes, the blurring can be modeled by a Gaussian point spread function

K(x, y, x′, y′

)= 1

2πσ σexp

(−1

2

(x − x′σ

)2

− 1

2

(y − y′σ

)2),

= ησ(x − x′)ησ (y − y′), (38)

where the function ησ is given by

ησ (t) = 1√2πσ

exp

(−1

2

(t

σ

)2),

and where σ and σ are two constants that characterize the blurring in the x and y direc-tions, respectively. If we introduce a Toeplitz matrix Tσ with elements given by

(Tσ )ij = 1√2πσ

exp

(−1

2

(i − jσ

)2), (39)

then the discretization of the atmospheric turbulence point spread function takes theform A ⊗ A = Tσ ⊗ Tσ , where the two Toeplitz matrices should have the appropriatedimensions. Depending on the dimensions and the value of σ , the matrix Tσ may be afull matrix or an effectively banded matrix. Figure 18 illustrates this point.

P.C. Hansen / Deconvolution and regularization 369

Figure 18. The decay of the function ησ (t) in the atmospheric turbulence blur model.

Figure 19. An exact image, and three blurred versions of the same image.

370 P.C. Hansen / Deconvolution and regularization

The above three blurring models are illustrated in figure 19 with an example usinga simple test image (which is generated by means of the function blur in REGULARIZA-TION TOOLS). Notice the distinctly different nature of atmospheric turbulence blur andout-of-focus blur.

We remark that in image processing, where the “signals” are always nonnegative,it is common [3] to measure the power of an image by first subtracting the image’s meanvalue from all pixels; i.e., if µF is the mean value of all the elements of F ,

µF = 1

mn

m∑i=1

n∑j=1

fij ,

then the power of F is given by

P(F) =m∑i=1

n∑j=1

(fij − µF )2 = ‖F − µFO‖2F,

in which O is an m × n matrix of all ones. Along this line, if Gpure = AFAT

is the“pure” blurred image and E is additive Gaussian noise with zero mean and standarddeviation σnoise, then the blurred signal-to-noise ratio (BSNR) of the blurred noisy imageG = Gpure + E is given by

BSNR = 10 log10P(Gpure)

P (E)= 10 log10

‖Gpure − µGpureO‖2F

mnσ 2noise

. (40)

Similarly, the improvement of the restored image Fres over the noisy blurred image G ismeasured by the improvement in SNR (ISNR), defined by

ISNR = 10 log10‖F −G‖2

F

‖F − Fres‖2F

. (41)

6.5. Regularization with Kronecker products

In this section we express the TSVD and Tikhonov solutions to the discretized 2-Dproblem (34) in terms of Kronecker products, thus illustrating how all computations withthe large n2 × n2 matrix A⊗ A can be avoided.

We begin with the TSVD solution which is defined in terms of the SVD of A⊗A.Let vec(Fk) denote the TSVD solution, and Fk the corresponding 2-D version. Here, kis the TSVD truncation parameter, i.e., the number of singular values of A⊗ A that areincluded in the regularized solution. Then vec(Fk) is given by

vec(Fk) =∑i,j

(uj ⊗ ui)Tvec(G)

σ jσi

(vj ⊗ vi

),

P.C. Hansen / Deconvolution and regularization 371

where the summation is over the k largest quantities σ jσi . Using the relation (33) wearrive at the expression

Fk =∑i,j

uTi Guj

σiσ jviv

Tj (42)

with the same summation as above. Note that we express the 2-D TSVD solution interms of rank-one matrices viv

Tj . Equation (42) is very convenient for computing 2-D

TSVD solutions in terms of the two SVDs of A and A. If the 2-D integral equationcollapses into a 1-D problem in the variables x and x′, then A = 1, i.e., a 1 × 1 matrixwith SVD u1 = σ 1 = v1 = 1, and (42) reduces to the standard expression (9) for the1-D TSVD solution.

We illustrate the 2-D TSVD algorithm with a 100× 75 image deblurring problem.The original image shows J. H. Wilkinson, the founder of rounding error analysis. Theblurring matrices A and A are generated by means of equation (39) with σ = 2.5 andthen truncated to a total bandwidth of 21 elements. Figure 20 shows the original im-

Figure 20. Top row: original and noisy blurred images of dimension 100 × 75. Middle and bottom rows:2-D TSVD solutions Fk ; the numbers above each image are k and the ISNR.

372 P.C. Hansen / Deconvolution and regularization

age F , the noisy blurred image G = AFAT + E, and eight reconstructed images Fk .

Above each plot is given k and the corresponding ISNR. Notice how the image qualityincreases with k until the influence of the noise E becomes visible.

Consider now the Tikhonov solution. The two underlying formulations of Tikho-nov regularization are now

min{∥∥(A⊗ A)vec(F )− vec(G)

∥∥22 + λ2

∥∥vec(F )∥∥2

2

}and

min{∥∥AFAT −G∥∥2

F + λ2‖F‖2F

},

and the associated least squares problem is

min

∥∥∥∥(A⊗AλI

)vec(F )−

(vec(G)

0

)∥∥∥∥2

, (43)

where I is the identity matrix of order n2. If vec(Fλ) denotes the Tikhonov solution,where Fλ is the corresponding 2-D version, then vec(Fλ) can also be expressed in termsof the SVD of A⊗ A. We recognize that the 2-D filter factors are given by

fij =σ 2i σ

2j

σ 2i σ

2j + λ2

, i, j = 1, . . . , n,

and the expression for the 2-D Tikhonov solution becomes

Fλ =n∑i=1

n∑j=1

fijuTi Guj

σiσjviv

Tj . (44)

Both algorithms described above require the computation of the SVD of the twomatrices A and A, and the algorithms are therefore of complexity O(n3). In the compu-tation of the 2-D Tikhonov solution, it is possible to replace one of the SVDs (of eitherA or A) with a bidiagonalization of the matrix, as described in [13]; the complexity ofthe algorithm is the same. It is not clear whether it is possible to derive an efficient 2-DTikhonov regularization algorithm based on two bidiagonalizations of A and A.

6.6. Direct methods for 2-D problems

Unfortunately, there is not much hope that we can develop efficient direct algo-rithms for the general 2-D Tikhonov problem that can exploit the Toeplitz structure,because the above least squares formulation (43) involving vec(F ) cannot be turned intoan equivalent linear problem that involves F itself. For example, the corresponding“normal equations” take the form of a generalized Sylvester equation

ATAFATA+ λ2F = ATGA

which is not a standard linear system of equations for computing F .

P.C. Hansen / Deconvolution and regularization 373

The situation is better when A and A are symmetric Toeplitz matrices, for then wecan use Franklin’s method (12) to arrive at the problem(

A⊗ A+ λ2I)vec(F ) = vec(G).

For this method, the filter factors that appear in (44) are given by

fij = σiσj

σiσj + λ2, i, j = 1, . . . , n.

Due to the Toeplitz structure of A and A, the coefficient matrix A⊗ A + λI is a blockToeplitz matrix with Toeplitz blocks (but it is not a Kronecker product), and the sys-tem can be solved efficiently by means of a block version of the Levinson algorithmfrom [25].

As an alternative to the Tikhonov and Franklin solutions, one can use the 1-Dalgorithms as building blocks for computing the solutions to the two problems(

ATA+ λ2I)F(A

TA+ λ2

I) = ATGA

and

(A+ λI) F (A+ λI) = G.The filter factors for these soltuions are, respectively,

fij =σ 2i σ

2j

(σ 2i + λ2)(σ 2

j + λ2), i, j = 1, . . . , n

and

fij = σiσj

(σi + λ)(σj + λ), i, j = 1, . . . , n.

They resemble those for Franklin’s method, but they decay in a different way.

6.7. Kronecker products in iterative methods

In the light of the discussion in the previous section, it is of particular interest toapply iterative methods to the discretized 2-D problem in (34). For example, one step ofLandweber’s method becomes

vec(F (k)

) = vec(F (k−1)

)+ ω(A⊗ A)T(vec(G)− (A⊗ A)vec

(F (k−1)

))and by means of (30) and (33) we can rewrite this formula in the following, more con-venient form

R(k−1) = G− AF(k−1)AT, F (k) = F (k−1) + ωATR(k−1)A,

where R(k−1) is the residual matrix corresponding to the iteration matrix F (k−1). We seethat the basic “building blocks” in this algorithm are the multiplications of the iterationmatrix F (k−1) and the residual matrix R(k−1) from the left and right by A and A.

374 P.C. Hansen / Deconvolution and regularization

Similarly, we can rewrite the 2-D version of the CGLS algorithm in terms of thematrices A and A. To do this, we use the fact that ‖vec(X)‖2 = ‖X‖F, and one step ofthe 2-D CGLS algorithm takes the form

αk = ‖ATR(k−1)A‖2

F

‖AD(k−1)AT‖2

F

,

F (k)=F (k−1) + αkD(k−1),

R(k)=R(k−1) − αkAD(k−1)AT,

βk = ‖ATR(k)A‖2F

‖ATR(k−1)A‖2F

,

D(k)=ATR(k)A+ βkD(k−1)

with starting conditions F (0) = initial image, R(0) = G−AF(0)AT, andD(0) = ATR(0)A.

The advantage of this formulation is that we work solely with the 2-dimensional arraysthat represent the underlying 2-D functions.

Figure 21 shows eight CGLS solutions F (k) to the same image deblurring problemas in figure 20. Initially the CGLS solutions F (k) are better approximations than thecorresponding TSVD solutions Fk. But as k increases the influence of finite precisionsets in, and after k = 10000 iterations the CGLS solution F (k) resembles the TSVDsolution with 3000 components.

Figure 21. Eight CGLS solutions F(k) to the 100 × 75 image deblurring problem. Above each image isshown the iteration count k and the ISNR.

P.C. Hansen / Deconvolution and regularization 375

6.8. 2-D FFTs and convolution

When both A and A are Toeplitz matrices, we can use the techniques from theprevious chapter to derive an FFT-based algorithm for computing the matrix productAXA, where X is any matrix that conforms with the other two matrices.

Our discussion takes its basis in the 2-D convolution of two general 2-D periodicdiscrete signals, represented by the two m× n matrices F and H . Then the convolutionof F and H is another 2-D periodic signal which is represented by the matrix G whoseelements for i = 0, . . . , m− 1 and j = 0, . . . , n− 1 are given by

gij =m−1∑k=0

n−1∑5=0

fk5h(i−k)modm,(j−5)mod n,

where the two subscripts of h are taken modulo m and n, respectively.The matrix G can be computed in less than 2m2n2 flops by means of the 2-D FFT.

We use the same notation as in the previous chapter, in which we write the DFT of ann-vector x as DFT(x) = Fnx, where Fn is a complex symmetric matrix. Then the 2-DFFT of an m× n matrix X is defined as follows

DFT2(X) = FmXFn. (45)

In this expression, we could add transposition to the complex symmetric matrix Fm inorder to make the expression appear like the previous 2-D expressions, but we shallrefrain from this. We also define the 2-D inverse DFT as IDFT2(X) = F−1

m XF−1n . We

can now express G as

G = F−1m

((FmFFn)� (FmHFn)

)F−1n = IDFT2

(DFT2(F )� DFT2(H)

).

The 2-D DFT (45) requires the computation of the ordinary DFT of each of then columns of X, followed by the computation of the ordinary DFT of each of them rowsof the matrix FmX. Ignoring that X is real (because the intermediate result FmX iscomplex), all the m + n DFTs in a single 2-D DFT can be computed by means of theFFT in n(5m log2m)+m(5n log2 n) = 5mn(log2m+ log2 n) flops. This is also the flopcount for the 2-D inverse DFT. Hence, provided that DFT2(H) has been precomputed,G can be computed in 10mn(log2m+ log2 n) flops, which simplifies to 20n2 log2 n flopswhen m = n.

The special case of convolution that we focus on in this chapter is when the ker-nel K in the integral equation (26) separates into a product of two functions as in (27).In connection with discrete data, this corresponds to a situation where the matrix H is arank-one matrix given by the outer product of two vectors v and w of length m and n,respectively,

H = vwT = v ⊗ wT.

The elements of H are thus hij = viwj and we see that H can be considered as asampling of the function κ(x)ω(y) on a 2-D grid.

376 P.C. Hansen / Deconvolution and regularization

To derive the corresponding expression for G we write its elements as

gij =m−1∑k=0

n−1∑5=0

fk5v(i−k)mod mw(j−5)mod n

=m−1∑k=0

(n−1∑5=0

fk5w(j−5)mod n

)v(i−k)modm =

m−1∑k=0

γkj v(i−k)mod m,

where we have introduced γkj = ∑n−15=0 fk5w(j−5)mod n. These quantities define a ma-

trix Q whose rows are convolutions of w with the rows of F , i.e., QT = CwF T ⇔ Q =FCT

w. Moreover, the columns of G are convolutions of v with the columns of Q, i.e.,G = CvQ. We have thus shown that G is given by

G = CvFCTw, (46)

where Cv and Cw are two circulant matrices defined by setting their first columns equalto v and w, respectively. The 2-D DFT of H is particularly easy to compute and repre-sent, since

DFT2(H) = FmHFn = FmvwTF Tn = (Fmv)(Fnw)T = DFT(v)DFT(w)T,

i.e., we merely need to compute and store the DFTs of the two vectors v and w. Thisleads to an efficient algorithm for computing the matrix product in (46):

G = IDFT2(DFT2(F )� (DFT(v)DFT(w)T

)). (47)

Note that the elementwise matrix multiplication should be computed without formingthe outer product of the two DFTs explicitly.

The above relations can now be used to efficiently compute the matrix product

AFAT

when A and A are Toeplitz matrices. The key idea is again to embed the twoToeplitz matrices in two larger circulant matrices C and C as described in the previouschapter, and similarly augment X with zero blocks such that the augmented matrix

X =(X 00 0

)conforms with C and C. Then the desired matrix product is the upper left submatrix in

Z = CXCT =(AXA

TZ12

Z21 Z22

),

while the other three submatrices are discarded. The multiplications with C and C couldbe computed by means of the fast algorithms from the previous chapter, by first workingon the columns of X and then on the rows of CX, but we can avoid the many DFT-IDFTpairs by instead using equation (47). The complete amount of work in this approachis approximately 2 · 5 · 2m2n

(log2(2m)+ log2(2n)

) � 40mn(log2m + log2 n) flops– which should be compared with the 2mn(m + n) flops in a standard matrix–matrix

P.C. Hansen / Deconvolution and regularization 377

Figure 22. Comparison of the flop count 4n3 for ordinary matrix multiplication with the flop count60n2 log2 n for the FFT-based algorithm.

multiplication algorithm for computing AXAT. If m = n then the two numbers are

80n2 log2 n and 4n3. And if special care is taken to exploit the real data and result, thenthe factor can be reduced to 60. Figure 22 compares the two flop counts.

References

[1] H.-M. Adorf, Hubble space telescope image restoration in its fourth year, Inverse Problems 11 (1995)639–653.

[2] C.T.H. Baker, The Numerical Treatment of Integral Equations (Clarendon Press, Oxford, 1977).[3] M.R. Banham and A.K. Katsaggelos, Digital image restoration, IEEE Signal Proc. Mag. 14 (1997)

24–41.[4] M. Bertero, P. Brianzi and E.R. Pike, Super-resolution in confocal scanning microscopy, Inverse

Problems 3 (1987) 195–212.[5] Å. Björck, Numerical Methods for Least Squares Problems (SIAM, Philadelphia, PA, 1996).[6] A.W. Bojanczyk, R.P. Brent and F.R. de Hoog, QR factorization of Toeplitz matrices, Numer. Math.

49 (1986) 210–221.[7] W.L. Briggs and V.E. Henson, The DFT. An Owner’s Manual for the Discrete Fourier Transform

(SIAM, Philadelphia, PA, 1995).[8] R.H. Chan, J.G. Nagy and R.J. Plemmons, FFT-based preconditioners for Toeplitz-block least squares

problems, SIAM J. Numer. Anal. 30 (1993) 1740–1768.[9] T. Chan, An optimal circulant preconditioner for Toeplitz systems, SIAM J. Sci. Statist. Comput. 9

(1988) 766–771.[10] T.F. Chan and P.C. Hansen, A look-ahead Levinson algorithm for general Toeplitz systems, IEEE

Trans. Signal Processing 40 (1992) 1079–1090.[11] L. Eldén, An efficient algorithm for the regularization of ill-conditioned least squares problems with

triangular Toeplitz matrices, SIAM J. Sci. Statist. Comput. 5 (1984) 229–236.

378 P.C. Hansen / Deconvolution and regularization

[12] L. Eldén, Numerical solution of the sideways heat equation, in: Inverse Problems in DiffusionProcesses, eds. H.W. Engl and W. Rundell (SIAM, Philadelphia, PA, 1995).

[13] L. Eldén and I. Skoglund, Algorithms for the regularization of ill-conditioned least squares problemswith tensor product structure, and applications to space-invariant image restoration, Report LiTH-MAT-R-1982-48, Department of Mathematics, Linköping University, Sweden (1982).

[14] J.N. Franklin, Minimum principles for ill-posed problems, SIAM J. Math. Anal. 9 (1978) 638–650.[15] G.H. Golub and C.F. Van Loan, Matrix Computations, 3rd ed. (Johns Hopkins Univ. Press, Baltimore,

MD, 1996).[16] M. Hanke and J.G. Nagy, Restoration of atmospherically blurred images by symmetric indefinite

conjugate gradient techniques, Inverse Problems 12 (1996) 157–173.[17] M. Hanke and C.R. Vogel, Two-level preconditioners for regularized inverse problems I: Theory,

Numer. Math. 83 (1999) 385–402.[18] P.C. Hansen, Truncated SVD solutions to discrete ill-posed problems with ill-determined numerical

rank, SIAM J. Sci. Statist. Comput. 11 (1990) 503–518.[19] P.C. Hansen, Regularization tools: A Matlab package for analysis and solution of discrete ill-posed

problems, Numer. Algorithms 6 (1994) 1–35.[20] P.C. Hansen, Rank-Deficient and Discrete Ill-Posed Problems. Numerical Aspects of Linear Inversion

(SIAM, Philadelphia, PA, 1998).[21] P.C. Hansen, Regularization tools version 3.0 for Matlab 5.2, Numer. Algorithms 20 (1999) 195–196;

software and manual is available from Netlib in directory numeralgo/na4.[22] M.R. Hestenes and E. Stiefel, Methods of conjugate gradients for solving linear systems, J. Res. Nat.

Bur. Standards 49 (1952) 409–436.[23] P.A. Jansson, ed. Deconvolution of Images and Spectra, 2nd ed. (Academic Press, San Diego, 1997).[24] T. Kailath and A.H. Sayed, eds., Fast Reliable Algorithms for Matrices with Structure (SIAM,

Philadelphia, PA, 1999).[25] N. Kalauptsidis, D. Manolakis and G. Carayannis, A family of computationally efficient algorithms

for multichannel signal processing – a tutorial review, Signal Processing 5 (1983) 5–19.[26] M. Kilmer and D.P. O’Leary, Pivoted Cauchy-like preconditioners for regularized solution of ill-posed

problems, SIAM J. Sci. Comput. 21 (1999) 88–110.[27] M. Kilmer and G.W. Stewart, Iterative regularization and MINRES, SIAM J. Matrix Anal. Appl. 21

(1999) 613–628.[28] J.G. Nagy, R.J. Plemmons and T.C. Torgersen, Iterative image restoration using approximate inverse

preconditioning, IEEE Trans. Image Processing 5 (1996) 1151–1162.[29] H. Park and L. Eldén, Stability analysis and fast algorithms for triangularization of Toeplitz matrices,

Numer. Math. 76 (1997) 383–402.[30] H. Park and L. Eldén, Schur-type methods for solving least squares problems with Toeplitz structure,

SIAM J. Sci. Comput. 22 (2000) 406–430.[31] D.L. Phillips, A technique for the numerical solution of certain integral equations of the first kind,

J. Assoc. Comput. Mach. 9 (1962) 84–97.[32] B.J. Sullivan and B. Liu, On the use of singular value decomposition and decimation in discrete-time

band-limited signal extrapolation, IEEE Trans. Acoust. Speech Signal Processing 32 (1984) 1201–1212.

[33] A.N. Tikhonov, Solution of incorrectly formulated problems and the regularization method, SovietMath. Dokl. 4 (1963) 1035–1038; English translation of Dokl. Akad. Nauk. SSSR 51 (1963) 501–504.

[34] C.F. Van Loan, Computational Frameworks for the Fast Fourier Transform (SIAM, Philadelphia, PA,1992).

[35] G.M. Wing and J.D. Zahrt, A Primer on Integral Equations of the First Kind (SIAM, Philadelphia,PA, 1991).

Recommended